Definition of the Opus Audio CodecOctasic Inc.4101, Molson StreetMontrealQuebecCanada+1 514 282-8858jean-marc.valin@octasic.comSkype Technologies S.A.Stadsgarden 6Stockholm11645SE+46 855 921 989koen.vos@skype.net
General
This document describes the Opus codec, designed for interactive speech and audio
transmission over the Internet.
We propose the Opus codec based on a linear prediction layer (LP) and an
MDCT-based layer. The main idea behind the proposal is that
the speech low frequencies are usually more efficiently coded using
linear prediction codecs (such as CELP variants), while the higher frequencies
are more efficiently coded in the transform domain (e.g. MDCT). For low
sampling rates, the MDCT layer is not useful and only the LP-based layer is
used. On the other hand, non-speech signals are not always adequately coded
using linear prediction, so for music only the MDCT-based layer is used.
In this proposed prototype, the LP layer is based on the
SILK codec
and the MDCT layer is based on the
CELT codec
.
This is a work in progress.The primary normative part of this specification is provided by the source
code part of the document. The codec contains significant amounts of fixed-point
arithmetic which must be performed exactly, including all rounding considerations,
and so any useful specification must make extensive use of domain-specific
symbolic language to adequately define these operations. Additionally, any
conflict between the symbolic representation and the included reference
implementation must be resolved. For the practical reasons of compatibility and
testability it would be advantageous to give the reference implementation to
have priority in any disagreement. The C language is also one of the most
widely understood human-readable symbolic representations for machine
behavior. For these reasons this RFC utilizes the reference implementation
as the sole symbolic representation of the codec.While the symbolic representation is unambiguous and complete it is not
always the easiest way to understand the codec's operation. For this reason
this document also describes significant parts of the codec in english and
takes the opportunity to explain the rational behind many of the more
surprising elements of the design. These descriptions are intended to be
accurate and informative but the limitations of common english sometimes
result in ambiguity, so it is intended that the reader will always read
them alongside the symbolic representation. Numerous references to the
implementation are provided for this purpose. The descriptions sometimes
differs in ordering, or through mathematical simplification, from the
reference wherever such deviation made an explanation easier to understand.
For example, the right shift and left shift operations in the reference
implementation are often described using division and multiplication in the text.
In general, the text is focused on the 'what' and 'why' while the symbolic
representation most clearly provides the 'how'.
In hybrid mode, each frame is coded first by the LP layer and then by the MDCT
layer. In the current prototype, the cutoff frequency is 8 kHz. In the MDCT
layer, all bands below 8 kHz are discarded, such that there is no coding
redundancy between the two layers. Also both layers use the same instance of
the range coder to encode the signal, which ensures that no "padding bits" are
wasted. The hybrid approach makes it easy to support both constant bit-rate
(CBR) and varaible bit-rate (VBR) coding. Although the SILK layer used is VBR,
it is easy to make the bit allocation of the CELT layer produce a final stream
that is CBR by using all the bits left unused by the SILK layer.
In addition to their frame size, the SILK and CELT codecs require
a look-ahead of 5.2 ms and 2.5 ms, respectively. SILK's look-ahead is due to
noise shaping estimation (5 ms) and the internal resampling (0.2 ms), while
CELT's look-ahead is due to the overlapping MDCT windows. To compensate for the
difference, the CELT encoder input is delayed by 2.7 ms. This ensures that low
frequencies and high frequencies arrive at the same time.
The source code is currently available in a
Git repository
which references two other
repositories (for SILK and CELT). Development snapshots are provided at
.
There are three possible operating modes for the proposed prototype:
A linear prediction (LP) mode for use in low bit-rate connections with up to 8 kHz audio bandwidth (16 kHz sampling rate)A hybrid (LP+MDCT) mode for full-bandwidth speech at medium bitratesAn MDCT-only mode for very low delay speech transmission as well as music transmission.
Each of these modes supports a number of difference frame sizes and sampling
rates. In order to distinguish between the various modes and configurations,
we define a single-byte table-of-contents (TOC) header that can used in the transport layer
(e.g RTP) to signal this information. The following describes the proposed
TOC byte.
The LP mode supports the following configurations (numbered from 0 to 11):
8 kHz: 10, 20, 40, 60 ms (0..3)12 kHz: 10, 20, 40, 60 ms (4..7)16 kHz: 10, 20, 40, 60 ms (8..11)
for a total of 12 configurations.
The hybrid mode supports the following configurations (numbered from 12 to 15):
32 kHz: 10, 20 ms (12..13)48 kHz: 10, 20 ms (14..15)
for a total of 4 configurations.
The MDCT-only mode supports the following configurations (numbered from 16 to 31):
8 kHz: 2.5, 5, 10, 20 ms (16..19)16 kHz: 2.5, 5, 10, 20 ms (20..23)32 kHz: 2.5, 5, 10, 20 ms (24..27)48 kHz: 2.5, 5, 10, 20 ms (28..31)
for a total of 16 configurations.
There is thus a total of 32 configurations, encoded in 5 bits. On bit is used to signal mono vs stereo, which leaves 2 bits for the number of frames per packets (codes 0 to 3):
0: 1 frames in the packet1: 2 frames in the packet, each with equal compressed size2: 2 frames in the packet, with different compressed size3: arbitrary number of frames in the packet
For code 2, the TOC byte is followed by the length of the first frame, encoded as described below.
For code 3, the TOC byte is followed by a byte encoding the number of frames in the packet, with the MSB indicating VBR. In the VBR case, the byte indicating the number of frames is followed by N-1 frame
lengths encoded as described below. As an additional limit, the audio duration contained
within a packet may not exceed 120 ms.
The compressed size of the frames (if needed) is indicated -- usually -- with one byte, with the following meaning:
0: No frame (DTX or lost packet)1-251: Size of the frame in bytes252-255: A second byte is needed. The total size is (size[1]*4)+size[0]
The maximum size representable is 255*4+255=1275 bytes. For 20 ms frames, that
represents a bit-rate of 510 kb/s, which is really the highest rate anyone would want
to use in stereo mode (beyond that point, lossless codecs would be more appropriate).
Simplest case: one narrowband mono 20-ms SILK frame
Two 48 kHz mono 5 ms CELT frames of the same compressed size:
Two 48 kHz mono 20-ms hybrid frames of different compressed size:
Four 48 kHz stereo 20-ms CELT frame of the same compressed size:
The Opus decoder consists of two main blocks: the SILK decoder and the CELT decoder.
The output of the Opus decode is the sum of the outputs from the SILK and CELT decoders
with proper sample rate conversion and delay compensation as illustrated in the
block diagram below. At any given time, one or both of the SILK and CELT decoders
may be active.
The range decoder extracts the symbols and integers encoded using the range encoder in
. The range decoder maintains an internal
state vector composed of the two-tuple (dif,rng), representing the
difference between the high end of the current range and the actual
coded value, and the size of the current range, respectively. Both
dif and rng are 32-bit unsigned integer values. rng is initialized to
2^7. dif is initialized to rng minus the top 7 bits of the first
input octet. Then the range is immediately normalized, using the
procedure described in the following section.
Decoding symbols is a two-step process. The first step determines
a value fs that lies within the range of some symbol in the current
context. The second step updates the range decoder state with the
three-tuple (fl,fh,ft) corresponding to that symbol, as defined in
.
The first step is implemented by ec_decode()
(rangedec.c),
and computes fs = ft-min((dif-1)/(rng/ft)+1,ft), where ft is
the sum of the frequency counts in the current context, as described
in . The divisions here are exact integer division.
In the reference implementation, a special version of ec_decode()
called ec_decode_bin() (rangeenc.c) is defined using
the parameter ftb instead of ft. It is mathematically equivalent to
calling ec_decode() with ft = (1<<ftb), but avoids one of the
divisions.
The decoder then identifies the symbol in the current context
corresponding to fs; i.e., the one whose three-tuple (fl,fh,ft)
satisfies fl <= fs < fh. This tuple is used to update the decoder
state according to dif = dif - (rng/ft)*(ft-fh), and if fl is greater
than zero, rng = (rng/ft)*(fh-fl), or otherwise rng = rng - (rng/ft)*(ft-fh). After this update, the range is normalized.
To normalize the range, the following process is repeated until
rng > 2^23. First, rng is set to (rng<8)&0xFFFFFFFF. Then the next
8 bits of input are read into sym, using the remaining bit from the
previous input octet as the high bit of sym, and the top 7 bits of the
next octet for the remaining bits of sym. If no more input octets
remain, zero bits are used instead. Then, dif is set to
(dif<<8)-sym&0xFFFFFFFF (i.e., using wrap-around if the subtraction
overflows a 32-bit register). Finally, if dif is larger than 2^31,
dif is then set to dif - 2^31. This process is carried out by
ec_dec_normalize() (rangedec.c).
Functions ec_dec_uint() or ec_dec_bits() are based on ec_decode() and
decode one of N equiprobable symbols, each with a frequency of 1,
where N may be as large as 2^32-1. Because ec_decode() is limited to
a total frequency of 2^16-1, this is done by decoding a series of
symbols in smaller contexts.
ec_dec_bits() (entdec.c) is defined, like
ec_decode_bin(), to take a single parameter ftb, with ftb < 32.
and ftb < 32, and produces an ftb-bit decoded integer value, t,
initialized to zero. While ftb is greater than 8, it decodes the next
8 most significant bits of the integer, s = ec_decode_bin(8), updates
the decoder state with the 3-tuple (s,s+1,256), adds those bits to
the current value of t, t = t<<8 | s, and subtracts 8 from ftb. Then
it decodes the remaining bits of the integer, s = ec_decode_bin(ftb),
updates the decoder state with the 3 tuple (s,s+1,1<<ftb), and adds
those bits to the final values of t, t = t<<ftb | s.
ec_dec_uint() (entdec.c) takes a single parameter,
ft, which is not necessarily a power of two, and returns an integer,
t, with a value between 0 and ft-1, inclusive, which is initialized to zero. Let
ftb be the location of the highest 1 bit in the two's-complement
representation of (ft-1), or -1 if no bits are set. If ftb>8, then
the top 8 bits of t are decoded using t = ec_decode((ft-1>>ftb-8)+1),
the decoder state is updated with the three-tuple
(s,s+1,(ft-1>>ftb-8)+1), and the remaining bits are decoded with
t = t<<ftb-8|ec_dec_bits(ftb-8). If, at this point, t >= ft, then
the current frame is corrupt, and decoding should stop. If the
original value of ftb was not greater than 8, then t is decoded with
t = ec_decode(ft), and the decoder state is updated with the
three-tuple (t,t+1,ft).
The bit allocation routines in CELT need to be able to determine a
conservative upper bound on the number of bits that have been used
to decode from the current frame thus far. This drives allocation
decisions which must match those made in the encoder. This is
computed in the reference implementation to fractional bit precision
by the function ec_dec_tell() (rangedec.c). Like all
operations in the range decoder, it must be implemented in a
bit-exact manner, and must produce exactly the same value returned by
ec_enc_tell() after encoding the same symbols.
At the receiving end, the received packets are by the range decoder split into a number of frames contained in the packet. Each of which contains the necessary information to reconstruct a 20 ms frame of the output signal.
An overview of the decoder is given in .
The range decoder decodes the encoded parameters from the received bitstream. Output from this function includes the pulses and gains for the excitation signal generation, as well as LTP and LSF codebook indices, which are needed for decoding LTP and LPC coefficients needed for LTP and LPC synthesis filtering the excitation signal, respectively.
Pulses and gains are decoded from the parameters that was decoded by the range decoder.
When a voiced frame is decoded and LTP codebook selection and indices are received, LTP coefficients are decoded using the selected codebook by choosing the vector that corresponds to the given codebook index in that codebook. This is done for each of the four subframes.
The LPC coefficients are decoded from the LSF codebook by first adding the chosen vectors, one vector from each stage of the codebook. The resulting LSF vector is stabilized using the same method that was used in the encoder, see
. The LSF coefficients are then converted to LPC coefficients, and passed on to the LPC synthesis filter.
The pulses signal is multiplied with the quantization gain to create the excitation signal.
For voiced speech, the excitation signal e(n) is input to an LTP synthesis filter that will recreate the long term correlation that was removed in the LTP analysis filter and generate an LPC excitation signal e_LPC(n), according to
using the pitch lag L, and the decoded LTP coefficients b_i.
For unvoiced speech, the output signal is simply a copy of the excitation signal, i.e., e_LPC(n) = e(n).
In a similar manner, the short-term correlation that was removed in the LPC analysis filter is recreated in the LPC synthesis filter. The LPC excitation signal e_LPC(n) is filtered using the LTP coefficients a_i, according to
where d_LPC is the LPC synthesis filter order, and y(n) is the decoded output signal.
Insert decoder figure.
Symbol(s)PDFConditionsilence[32767, 1]/32768post-filter[1, 1]/2octaveuniform (6)post-filterperiodraw bits (4+octave)post-filtergainraw bits (3)post-filtertapset[2, 1, 1]/4post-filtertransient[7, 1]/8intra[7, 1]/8coarse energytf_changetf_select[1, 1]/2spread[7, 2, 21, 2]/32dyn. alloc.alloc. trim[2, 2, 5, 10, 22, 46, 22, 10, 5, 2, 2]/128skip[1, 1]/2intensityuniformdual[1, 1]/2fine energyresidualanti-collapse[1, 1]/2finalizeOrder of the symbols in the CELT section of the bit-stream.
The decoder extracts information from the range-coded bit-stream in the order
described in the figure above. In some circumstances, it is
possible for a decoded value to be out of range due to a very small amount of redundancy
in the encoding of large integers by the range coder.
In that case, the decoder should assume there has been an error in the coding,
decoding, or transmission and SHOULD take measures to conceal the error and/or report
to the application that a problem has occurred.
The transient flag encoded in the bit-stream has a
probability of 1/8. When it is set, then the MDCT coefficients represent multiple
short MDCTs in the frame. When not set, the coefficients represent a single
long MDCT for the frame. In addition to the global transient flag is a per-band
binary flag to change the time-frequency (tf) resolution independently in each band. The
change in tf resolution is defined in tf_select_table[][] in celt.c and depends
on the frame size, whether the transient flag is set, and the value of tf_select.
The tf_select flag uses a 1/2 probability, but is only decoded
if it can have an impact on the result knowing the value of all per-band
tf_change flags.
It is important to quantize the energy with sufficient resolution because
any energy quantization error cannot be compensated for at a later
stage. Regardless of the resolution used for encoding the shape of a band,
it is perceptually important to preserve the energy in each band. CELT uses a
three-step coarse-fine-fine strategy for encoding the energy in the base-2 log
domain, as implemented in quant_bands.c
Coarse quantization of the energy uses a fixed resolution of 6 dB
(integer part of base-2 log). To minimize the bitrate, prediction is applied
both in time (using the previous frame) and in frequency (using the previous
bands). The part of the prediction that is based on the
previous frame can be disabled, creating an "intra" frame where the energy
is coded without reference to prior frames. The decoder first reads the intra flag
to determine what prediction is used.
The 2-D z-transform of
the prediction filter is: A(z_l, z_b)=(1-a*z_l^-1)*(1-z_b^-1)/(1-b*z_b^-1)
where b is the band index and l is the frame index. The prediction coefficients
applied depend on the frame size in use when not using intra energy and a=0 b=4915/32768
when using intra energy.
The time-domain prediction is based on the final fine quantization of the previous
frame, while the frequency domain (within the current frame) prediction is based
on coarse quantization only (because the fine quantization has not been computed
yet). The prediction is clamped internally so that fixed point implementations with
limited dynamic range to not suffer desynchronization.
We approximate the ideal
probability distribution of the prediction error using a Laplace distribution
with seperate parameters for each frame size in intra and inter-frame modes. The
coarse energy quantization is performed by unquant_coarse_energy() and
unquant_coarse_energy_impl() (quant_bands.c). The encoding of the Laplace-distributed values is
implemented in ec_laplace_decode() (laplace.c).
The number of bits assigned to fine energy quantization in each band is determined
by the bit allocation computation described in .
Let B_i be the number of fine energy bits
for band i; the refinement is an integer f in the range [0,2^B_i-1]. The mapping between f
and the correction applied to the coarse energy is equal to (f+1/2)/2^B_i - 1/2. Fine
energy quantization is implemented in quant_fine_energy() (quant_bands.c).
When some bits are left "unused" after all other flags have been decoded, these bits
are assigned to a "final" step of fine allocation. In effect, these bits are used
to add one extra fine energy bit per band per channel. The allocation process
determines two priorities for the final fine bits.
Any remaining bits are first assigned only to bands of priority 0, starting
from band 0 and going up. If all bands of priority 0 have received one bit per
channel, then bands of priority 1 are assigned an extra bit per channel,
starting from band 0. If any bit is left after this, they are left unused.
This is implemented in unquant_energy_finalise() (quant_bands.c).
Many codecs transmit significant amounts of side information for
the purpose of controlling bit allocation within a frame. Often this
side information controls bit usage indirectly and must be carefully
selected to achieve the desired rate constraints.The band-energy normalized structure of Opus MDCT mode ensures that a
constant bit allocation for the shape content of a band will result in a
roughly constant tone to noise ratio, which provides for fairly consistent
perceptual performance. The effectiveness of this approach is the result of
two factors: The band energy, which is understood to be perceptually
important on its own, is always preserved regardless of the shape precision and because
the constant tone-to-noise ratio implies a constant intra-band noise to masking ratio.
Intra-band masking is the strongest of the perceptual masking effects. This structure
means that the ideal allocation is more consistent from frame to frame than
it is for other codecs without an equivalent structure.Because the bit allocation is used to drive the decoding of the range-coder
stream it MUST be recovered exactly so that identical coding decisions are
made in the encoder and decoder. Any deviation from the reference's resulting
bit allocation will result in corrupted output, though implementers are
free to implement the procedure in any way which produces identical results.Because all of the information required to decode a frame must be derived
from that frame alone in order to retain robustness to packet loss the
overhead of explicitly signaling the allocation would be considerable,
especially for low-latency (small frame size) applications,
even though the allocation is relatively static.For this reason, in the MDCT mode Opus uses a primarily implicit bit
allocation. The available bit-stream capacity is known in advance to both
the encoder and decoder without additional signaling, ultimately from the
packet sizes expressed by a higher level protocol. Using this information
the codec interpolates an allocation from a hard-coded table.While the band-energy structure effectively models intra-band masking,
it ignores the weaker inter-band masking, band-temporal masking, and
other less significant perceptual effects. While these effects can
often be ignored they can become significant for particular samples. One
mechanism available to encoders would be to simply increase the overall
rate for these frames, but this is not possible in a constant rate mode
and can be fairly inefficient. As a result three explicitly signaled
mechanisms are provided to alter the implicit allocation:Band boostAllocation trimband skippingThe first of these mechanisms, band boost, allows an encoder to boost
the allocation in specific bands. The second, allocation trim, works by
biasing the overall allocation towards higher or lower frequency bands. The third, band
skipping, selects which low-precision high frequency bands
will be allocated no shape bits at all.In stereo mode there are also two additional parameters
potentially coded as part of the allocation procedure: a parameter to allow the
selective elimination of allocation for the 'side' in jointly coded bands,
and a flag to deactivate joint coding. These values are not signaled if
they would be meaningless in the overall context of the allocation.Because every signaled adjustment increases overhead and implementation
complexity none were included speculatively: The reference encoder makes use
of all of these mechanisms. While the decision logic in the reference was
found to be effective enough to justify the overhead and complexity further
analysis techniques may be discovered which increase the effectiveness of these
parameters. As with other signaled parameters, encoder is free to choose the
values in any manner but unless a technique is known to deliver superior
perceptual results the methods used by the reference implementation should be
used.The process of allocation consists of the following steps: determining the per-band
maximum allocation vector, decoding the boosts, decoding the tilt, determining
the remaining capacity the frame, searching the mode table for the
entry nearest but not exceeding the available space (subject to the tilt, boosts, band
maximums, and band minimums), linear interpolation, reallocation of
unused bits with concurrent skip decoding, determination of the
fine-energy vs shape split, and final reallocation. This process results
in an shape allocation per-band (in 1/8th bit units), a per-band fine-energy
allocation (in 1 bit per channel units), a set of band priorities for
controlling the use of remaining bits at the end of the frame, and a
remaining balance of unallocated space which is usually zero except
at very high rates.The maximum allocation vector is an approximation of the maximum space
which can be used by each band for a given mode. The value is
approximate because the shape encoding is variable rate (due
to entropy coding of splitting parameters). Setting the maximum too low reduces the
maximum achievable quality in a band while setting it too high
may result in waste: bit-stream capacity available at the end
of the frame which can not be put to any use. The maximums
specified by the codec reflect the average maximum. In the reference
the maximums are provided partially computed form, in order to fit in less
memory, as a static table (XXX cache.caps). Implementations are expected
to simply use the same table data but the procedure for generating
this table is included in rate.c as part of compute_pulse_cache().To convert the values in cache.caps into the actual maximums: First
set nbBands to the maximum number of bands for this mode and stereo to
zero if stereo is not in use and one otherwise. For each band assign N
to the number of MDCT bins covered by the band (for one channel), set LM
to the shift value for the frame size (e.g. 0 for 120, 1 for 240, 3 for 480)
then set i to nbBands*(2*LM+stereo). Then set the maximum for the band to
the i-th index of cache.caps + 64 and multiply by the number of channels
in the current frame (one or two) and by N then divide the result by 4
using truncating integer division. The resulting vector will be called
cap[]. The elements fit in signed 16 bit integers but do not fit in 8 bits.
This procedure is implemented in the reference in the function init_caps() in celt.c.
The band boosts are represented by a series of binary symbols which
are coded with very low probability. Each band can potentially be boosted
multiple times, subject to the frame actually having enough room to obey
the boost and having enough room to code the boost symbol. The default
coding cost for a boost starts out at six bits, but subsequent boosts
in a band cost only a single bit and every time a band is boosted the
initial cost is reduced (down to a minimum of two). Since the initial
cost of coding a boost is 6 bits the coding cost of the boost symbols when
completely unused is 0.48 bits/frame for a 21 band mode (21*-log2(1-1/2^6)).To decode the band boosts: First set 'dynalloc_logp' to 6, the initial
amount of storage required to signal a boost in bits, 'total_bits' to the
size of the frame in 8th-bits, 'total_boost' to zero, and 'tell' to the total number
of 8th bits decoded
so far. For each band from the coding start (0 normally, but 17 in hybrid mode)
to the coding end (which changes depending on the signaled bandwidth): Set 'width'
to the number of MDCT bins in this band for all channels. Take the larger of width
and 64, then the minimum of that value and the width times eight and set 'quanta'
to the result. This represents a boost step size of six bits subject to limits
of 1/bit/sample and 1/8th bit/sample. Set 'boost' to zero and 'dynalloc_loop_logp'
to dynalloc_logp. While dynalloc_loop_log (the current worst case symbol cost) in
8th bits plus tell is less than total_bits plus total_boost and boost is less than cap[] for this
band: Decode a bit from the bitstream with a with dynalloc_loop_logp as the cost
of a one, update tell to reflect the current used capacity, if the decoded value
is zero break the loop otherwise add quanta to boost and total_boost, subtract quanta from
total_bits, and set dynalloc_loop_log to 1. When the while loop finishes
boost contains the boost for this band. If boost is non-zero and dynalloc_logp
is greater than 2 decrease dynalloc_logp. Once this process has been
execute on all bands the band boosts have been decoded. This procedure
is implemented around line 2352 of celt.c.At very low rates it's possible that there won't be enough available
space to execute the inner loop even once. In these cases band boost
is not possible but its overhead is completely eliminated. Because of the
high cost of band boost when activated a reasonable encoder should not be
using it at very low rates. The reference implements its dynalloc decision
logic at around 1269 of celt.cThe allocation trim is a integer value from 0-10. The default value of
5 indicates no trim. The trim parameter is entropy coded in order to
lower the coding cost of less extreme adjustments. Values lower than
5 bias the allocation towards lower frequencies and values above 5
bias it towards higher frequencies. Like other signaled parameters, signaling
of the trim is gated so that it is not included if there is insufficient space
available in the bitstream. To decode the trim first set
the trim value to 5 then iff the count of decoded 8th bits so far (ec_tell_frac)
plus 48 (6 bits) is less than or equal to the total frame size in 8th
bits minus total_boost (a product of the above band boost procedure) then
decode the trim value using the inverse CDF {127, 126, 124, 119, 109, 87, 41, 19, 9, 4, 2, 0}.Stereo parametersAnti-collapse reservationThe allocation computation first begins by setting up some initial conditions.
'total' is set to the available remaining 8th bits, computed by taking the
size of the coded frame times 8 and subtracting ec_tell_frac(). From this value one (8th bit)
is subtracted to assure that the resulting allocation will be conservative. 'anti_collapse_rsv'
is set to 8 (8th bits) iff the frame is a transient, LM is greater than 1, and total is
greater than or equal to (LM+2) * 8. Total is then decremented by anti_collapse_rsv and clamped
to be equal to or greater than zero. 'skip_rsv' is set to 8 (8th bits) if total is greater than
8, otherwise it is zero. Total is then decremented by skip_rsv. This reserves space for the
final skipping flag.If the current frame is stereo intensity_rsv is set to the conservative log2 in 8th bits
of the number of coded bands for this frame (given by the table LOG2_FRAC_TABLE). If
intensity_rsv is greater than total then intensity_rsv is set to zero otherwise total is
decremented by intensity_rsv, and if total is still greater than 8 dual_stereo_rsv is
set to 8 and total is decremented by dual_stereo_rsv.The allocation process then computes a vector representing the hard minimum amounts allocation
any band will receive for shape. This minimum is higher than the technical limit of the PVQ
process, but very low rate allocations produce excessively an sparse spectrum and these bands
are better served by having no allocation at all. For each coded band set thresh[band] to
twenty-four times the number of MDCT bins in the band and divide by 16. If 8 times the number
of channels is greater, use that instead. This sets the minimum allocation to one bit per channel
or 48 128th bits per MDCT bin, whichever is greater. The band size dependent part of this
value is not scaled by the channel count because at the very low rates where this limit is
applicable there will usually be no bits allocated to the side.The previously decoded allocation trim is used to derive a vector of per-band adjustments,
'trim_offsets[]'. For each coded band take the alloc_trim and subtract 5 and LM then multiply
the result by number of channels, the number MDCT bins in the shortest frame size for this mode,
the number remaining bands, 2^LM, and 8. Then divide this value by 64. Finally, if the
number of MDCT bins in the band per channel is only one 8 times the number of channels is subtracted
in order to diminish the allocation by one bit because width 1 bands receive greater benefit
from the coarse energy coding.
In each band, the normalized shape is encoded
using a vector quantization scheme called a "Pyramid vector quantizer".
In
the simplest case, the number of bits allocated in
is converted to a number of pulses as described
by . Knowing the number of pulses and the
number of samples in the band, the decoder calculates the size of the codebook
as detailed in . The size is used to decode
an unsigned integer (uniform probability model), which is the codeword index.
This index is converted into the corresponding vector as explained in
. This vector is then scaled to unit norm.
Although the allocation is performed in 1/8th bit units, the quantization requires
an integer number of pulses K. To do this, the encoder searches for the value
of K that produces the number of bits that is the nearest to the allocated value
(rounding down if exactly half-way between two values), subject to not exceeding
the total number of bits available. For efficiency reasons the search is performed against a
precomputated allocation table which only permits some K values for each N. The number of
codebooks entries can be computed as explained in . The difference
between the number of bits allocated and the number of bits used is accumulated to a
balance (initialised to zero) that helps adjusting the
allocation for the next bands. One third of the balance is applied to the
bit allocation of the each band to help achieving the target allocation. The only
exceptions are the band before the last and the last band, for which half the balance
and the whole balance are applied, respectively.
The codeword is decoded as a uniformly-distributed integer value
by decode_pulses() (cwrs.c).
The codeword is converted from a unique index in the same way as specified in
. The indexing is based on the calculation of V(N,K)
(denoted N(L,K) in ), which is the number of possible
combinations of K pulses
in N samples. The number of combinations can be computed recursively as
V(N,K) = V(N-1,K) + V(N,K-1) + V(N-1,K-1), with V(N,0) = 1 and V(0,K) = 0, K != 0.
There are many different ways to compute V(N,K), including pre-computed tables and direct
use of the recursive formulation. The reference implementation applies the recursive
formulation one line (or column) at a time to save on memory use,
along with an alternate,
univariate recurrence to initialise an arbitrary line, and direct
polynomial solutions for small N. All of these methods are
equivalent, and have different trade-offs in speed, memory usage, and
code size. Implementations MAY use any methods they like, as long as
they are equivalent to the mathematical definition.
The decoding of the codeword from the index is performed as specified in
, as implemented in function
decode_pulses() (cwrs.c).
To avoid the need for multi-precision calculations when decoding PVQ codevectors,
the maximum size allowed for codebooks is 32 bits. When larger codebooks are
needed, the vector is instead split in two sub-vectors of size N/2.
A quantized gain parameter with precision
derived from the current allocation is entropy coded to represent the relative
gains of each side of the split and the entire decoding process is recursively
applied. Multiple levels of splitting may be applied up to a frame size
dependent limit. The same recursive mechanism is applied for the joint coding
of stereo audio.
When the frame has the transient bit set...
Just like each band was normalized in the encoder, the last step of the decoder before
the inverse MDCT is to denormalize the bands. Each decoded normalized band is
multiplied by the square root of the decoded energy. This is done by denormalise_bands()
(bands.c).
The inverse MDCT implementation has no special characteristics. The
input is N frequency-domain samples and the output is 2*N time-domain
samples, while scaling by 1/2. The output is windowed using the same window
as the encoder. The IMDCT and windowing are performed by mdct_backward
(mdct.c). If a time-domain pre-emphasis
window was applied in the encoder, the (inverse) time-domain de-emphasis window
is applied on the IMDCT result.
The output of the inverse MDCT (after weighted overlap-add) is sent to the
post-filter. Although the post-filter is applied at the end, the post-filter
parameters are encoded at the beginning, just after the silence flag.
The post-filter can be switched on or off using one bit (logp=1).
If the post-filter is enabled, then the octave is decoded as an integer value
between 0 and 6 of uniform probability. Once the octave is known, the fine pitch
within the octave is decoded using 4+octave raw bits. The final pitch period
is equal to (16<<octave)+fine_pitch-1 so it is bounded between 15 and 1022,
inclusively. Next, the gain is decoded as three raw bits and is equal to
G=3*(int_gain+1)/32. The set of post-filter taps is decoded last using
a pdf equal to [2, 1, 1]/4. Tapset zero corresponds to the filter coefficients
g0 = 0.3066406250, g1 = 0.2170410156, g2 = 0.1296386719. Tapset one
corresponds to the filter coefficients g0 = 0.4638671875, g1 = 0.2680664062,
g2 = 0, and tapset two uses filter coefficients g0 = 0.7998046875,
g1 = 0.1000976562, g2 = 0.
The post-filter response is thus computed as:
During a transition between different gains, a smooth transition is calculated
using the square of the MDCT window. It is important that values of y(n) be
interpolated one at a time such that the past value of y(n) used is interpolated.
After the post-filter,
the signal is de-emphasized using the inverse of the pre-emphasis filter
used in the encoder: 1/A(z)=1/(1-alpha_p*z^-1), where alpha_p=0.8500061035.
Packet loss concealment (PLC) is an optional decoder-side feature which
SHOULD be included when transmitting over an unreliable channel. Because
PLC is not part of the bit-stream, there are several possible ways to
implement PLC with different complexity/quality trade-offs. The PLC in
the reference implementation finds a periodicity in the decoded
signal and repeats the windowed waveform using the pitch offset. The windowed
waveform is overlapped in such a way as to preserve the time-domain aliasing
cancellation with the previous frame and the next frame. This is implemented
in celt_decode_lost() (mdct.c).
Opus encoder block diagram.
Opus uses an entropy coder based upon ,
which is itself a rediscovery of the FIFO arithmetic code introduced by .
It is very similar to arithmetic encoding, except that encoding is done with
digits in any base instead of with bits,
so it is faster when using larger bases (i.e.: an octet). All of the
calculations in the range coder must use bit-exact integer arithmetic.
The range coder also acts as the bit-packer for Opus. It is
used in three different ways, to encode:
entropy-coded symbols with a fixed probability model using ec_encode(), (rangeenc.c)integers from 0 to 2^M-1 using ec_enc_uint() or ec_enc_bits(), (entenc.c)integers from 0 to N-1 (where N is not a power of two) using ec_enc_uint(). (entenc.c)
The range encoder maintains an internal state vector composed of the
four-tuple (low,rng,rem,ext), representing the low end of the current
range, the size of the current range, a single buffered output octet,
and a count of additional carry-propagating output octets. Both rng
and low are 32-bit unsigned integer values, rem is an octet value or
the special value -1, and ext is an integer with at least 16 bits.
This state vector is initialized at the start of each each frame to
the value (0,2^31,-1,0).
Each symbol is drawn from a finite alphabet and coded in a separate
context which describes the size of the alphabet and the relative
frequency of each symbol in that alphabet. Opus only uses static
contexts; they are not adapted to the statistics of the data that is
coded.
The main encoding function is ec_encode() (rangeenc.c),
which takes as an argument a three-tuple (fl,fh,ft)
describing the range of the symbol to be encoded in the current
context, with 0 <= fl < fh <= ft <= 65535. The values of this tuple
are derived from the probability model for the symbol. Let f(i) be
the frequency of the ith symbol in the current context. Then the
three-tuple corresponding to the kth symbol is given by
ec_encode() updates the state of the encoder as follows. If fl is
greater than zero, then low = low + rng - (rng/ft)*(ft-fl) and
rng = (rng/ft)*(fh-fl). Otherwise, low is unchanged and
rng = rng - (rng/ft)*(fh-fl). The divisions here are exact integer
division. After this update, the range is normalized.
To normalize the range, the following process is repeated until
rng > 2^23. First, the top 9 bits of low, (low>>23), are placed into
a carry buffer. Then, low is set to . This process is carried out by
ec_enc_normalize() (rangeenc.c).
The 9 bits produced in each iteration of the normalization loop
consist of 8 data bits and a carry flag. The final value of the
output bits is not determined until carry propagation is accounted
for. Therefore the reference implementation buffers a single
(non-propagating) output octet and keeps a count of additional
propagating (0xFF) output octets. An implementation MAY choose to use
any mathematically equivalent scheme to perform carry propagation.
The function ec_enc_carry_out() (rangeenc.c) performs
this buffering. It takes a 9-bit input value, c, from the normalization
8-bit output and a carry bit. If c is 0xFF, then ext is incremented
and no octets are output. Otherwise, if rem is not the special value
-1, then the octet (rem+(c>>8)) is output. Then ext octets are output
with the value 0 if the carry bit is set, or 0xFF if it is not, and
rem is set to the lower 8 bits of c. After this, ext is set to zero.
In the reference implementation, a special version of ec_encode()
called ec_encode_bin() (rangeenc.c) is defined to
take a two-tuple (fl,ftb), where , but avoids using division.
Functions ec_enc_uint() or ec_enc_bits() are based on ec_encode() and
encode one of N equiprobable symbols, each with a frequency of 1,
where N may be as large as 2^32-1. Because ec_encode() is limited to
a total frequency of 2^16-1, this is done by encoding a series of
symbols in smaller contexts.
ec_enc_bits() (entenc.c) is defined, like
ec_encode_bin(), to take a two-tuple (fl,ftb), with >ftb-8&0xFF) using ec_encode_bin() and
subtracts 8 from ftb. Then, it encodes the remaining bits of fl, e.g.,
(fl&(1<, again using ec_encode_bin().
ec_enc_uint() (entenc.c) takes a two-tuple (fl,ft),
where ft is not necessarily a power of two. Let ftb be the location
of the highest 1 bit in the two's-complement representation of
(ft-1), or -1 if no bits are set. If ftb>8, then the top 8 bits of fl
are encoded using ec_encode() with the three-tuple
(fl>>ftb-8,(fl>>ftb-8)+1,(ft-1>>ftb-8)+1), and the remaining bits
are encoded with ec_enc_bits using the two-tuple
.
After all symbols are encoded, the stream must be finalized by
outputting a value inside the current range. Let end be the integer
in the interval [low,low+rng) with the largest number of trailing
zero bits. Then while end is not zero, the top 9 bits of end, e.g.,
>23), are sent to the carry buffer, and end is replaced by
(end<<8&0x7FFFFFFF). Finally, if the value in carry buffer, rem, is]]>
neither zero nor the special value -1, or the carry count, ext, is
greater than zero, then 9 zero bits are sent to the carry buffer.
After the carry buffer is finished outputting octets, the rest of the
output buffer is padded with zero octets. Finally, rem is set to the
special value -1. This process is implemented by ec_enc_done()
(rangeenc.c).
The bit allocation routines in Opus need to be able to determine a
conservative upper bound on the number of bits that have been used
to encode the current frame thus far. This drives allocation
decisions and ensures that the range code will not overflow the
output buffer. This is computed in the reference implementation to
fractional bit precision by the function ec_enc_tell()
(rangeenc.c).
Like all operations in the range encoder, it must
be implemented in a bit-exact manner.
In the following, we focus on the core encoder and describe its components. For simplicity, we will refer to the core encoder simply as the encoder in the remainder of this document. An overview of the encoder is given in .
The input signal is processed by a VAD (Voice Activity Detector) to produce a measure of voice activity, and also spectral tilt and signal-to-noise estimates, for each frame. The VAD uses a sequence of half-band filterbanks to split the signal in four subbands: 0 - Fs/16, Fs/16 - Fs/8, Fs/8 - Fs/4, and Fs/4 - Fs/2, where Fs is the sampling frequency, that is, 8, 12, 16 or 24 kHz. The lowest subband, from 0 - Fs/16 is high-pass filtered with a first-order MA (Moving Average) filter (with transfer function H(z) = 1-z^(-1)) to reduce the energy at the lowest frequencies. For each frame, the signal energy per subband is computed. In each subband, a noise level estimator tracks the background noise level and an SNR (Signal-to-Noise Ratio) value is computed as the logarithm of the ratio of energy to noise level. Using these intermediate variables, the following parameters are calculated for use in other SILK modules:
Average SNR. The average of the subband SNR values.
Smoothed subband SNRs. Temporally smoothed subband SNR values.
Speech activity level. Based on the average SNR and a weighted average of the subband energies.
Spectral tilt. A weighted average of the subband SNRs, with positive weights for the low subbands and negative weights for the high subbands.
The input signal is filtered by a high-pass filter to remove the lowest part of the spectrum that contains little speech energy and may contain background noise. This is a second order ARMA (Auto Regressive Moving Average) filter with a cut-off frequency around 70 Hz.
In the future, a music detector may also be used to lower the cut-off frequency when the input signal is detected to be music rather than speech.
The high-passed input signal is processed by the open loop pitch estimator shown in .
The pitch analysis finds a binary voiced/unvoiced classification, and, for frames classified as voiced, four pitch lags per frame - one for each 5 ms subframe - and a pitch correlation indicating the periodicity of the signal. The input is first whitened using a Linear Prediction (LP) whitening filter, where the coefficients are computed through standard Linear Prediction Coding (LPC) analysis. The order of the whitening filter is 16 for best results, but is reduced to 12 for medium complexity and 8 for low complexity modes. The whitened signal is analyzed to find pitch lags for which the time correlation is high. The analysis consists of three stages for reducing the complexity:
In the first stage, the whitened signal is downsampled to 4 kHz (from 8 kHz) and the current frame is correlated to a signal delayed by a range of lags, starting from a shortest lag corresponding to 500 Hz, to a longest lag corresponding to 56 Hz.
The second stage operates on a 8 kHz signal ( downsampled from 12, 16 or 24 kHz ) and measures time correlations only near the lags corresponding to those that had sufficiently high correlations in the first stage. The resulting correlations are adjusted for a small bias towards short lags to avoid ending up with a multiple of the true pitch lag. The highest adjusted correlation is compared to a threshold depending on:
Whether the previous frame was classified as voiced
The speech activity level
The spectral tilt.
If the threshold is exceeded, the current frame is classified as voiced and the lag with the highest adjusted correlation is stored for a final pitch analysis of the highest precision in the third stage.
The last stage operates directly on the whitened input signal to compute time correlations for each of the four subframes independently in a narrow range around the lag with highest correlation from the second stage.
The noise shaping analysis finds gains and filter coefficients used in the prefilter and noise shaping quantizer. These parameters are chosen such that they will fulfil several requirements:
Balancing quantization noise and bitrate. The quantization gains determine the step size between reconstruction levels of the excitation signal. Therefore, increasing the quantization gain amplifies quantization noise, but also reduces the bitrate by lowering the entropy of the quantization indices.Spectral shaping of the quantization noise; the noise shaping quantizer is capable of reducing quantization noise in some parts of the spectrum at the cost of increased noise in other parts without substantially changing the bitrate. By shaping the noise such that it follows the signal spectrum, it becomes less audible. In practice, best results are obtained by making the shape of the noise spectrum slightly flatter than the signal spectrum.Deemphasizing spectral valleys; by using different coefficients in the analysis and synthesis part of the prefilter and noise shaping quantizer, the levels of the spectral valleys can be decreased relative to the levels of the spectral peaks such as speech formants and harmonics. This reduces the entropy of the signal, which is the difference between the coded signal and the quantization noise, thus lowering the bitrate.Matching the levels of the decoded speech formants to the levels of the original speech formants; an adjustment gain and a first order tilt coefficient are computed to compensate for the effect of the noise shaping quantization on the level and spectral tilt. shows an example of an input signal spectrum (1). After de-emphasis and level matching, the spectrum has deeper valleys (2). The quantization noise spectrum (3) more or less follows the input signal spectrum, while having slightly less pronounced peaks. The entropy, which provides a lower bound on the bitrate for encoding the excitation signal, is proportional to the area between the deemphasized spectrum (2) and the quantization noise spectrum (3). Without de-emphasis, the entropy is proportional to the area between input spectrum (1) and quantization noise (3) - clearly higher.
The transformation from input signal to deemphasized signal can be described as a filtering operation with a filter
having an adjustment gain G, a first order tilt adjustment filter with
tilt coefficient c_tilt, and where
is the analysis part of the de-emphasis filter, consisting of the short-term shaping filter with coefficients a_ana(k), and the long-term shaping filter with coefficients b_ana(k) and pitch lag L. The parameter d determines the number of long-term shaping filter taps.
Similarly, but without the tilt adjustment, the synthesis part can be written as
All noise shaping parameters are computed and applied per subframe of 5 milliseconds. First, an LPC analysis is performed on a windowed signal block of 15 milliseconds. The signal block has a look-ahead of 5 milliseconds relative to the current subframe, and the window is an asymmetric sine window. The LPC analysis is done with the autocorrelation method, with an order of 16 for best quality or 12 in low complexity operation. The quantization gain is found as the square-root of the residual energy from the LPC analysis, multiplied by a value inversely proportional to the coding quality control parameter and the pitch correlation.
Next we find the two sets of short-term noise shaping coefficients a_ana(k) and a_syn(k), by applying different amounts of bandwidth expansion to the coefficients found in the LPC analysis. This bandwidth expansion moves the roots of the LPC polynomial towards the origo, using the formulas
where a(k) is the k'th LPC coefficient and the bandwidth expansion factors g_ana and g_syn are calculated as
where C is the coding quality control parameter between 0 and 1. Applying more bandwidth expansion to the analysis part than to the synthesis part gives the desired de-emphasis of spectral valleys in between formants.
The long-term shaping is applied only during voiced frames. It uses three filter taps, described by
For unvoiced frames these coefficients are set to 0. The multiplication factors F_ana and F_syn are chosen between 0 and 1, depending on the coding quality control parameter, as well as the calculated pitch correlation and smoothed subband SNR of the lowest subband. By having F_ana less than F_syn, the pitch harmonics are emphasized relative to the valleys in between the harmonics.
The tilt coefficient c_tilt is for unvoiced frames chosen as
for voiced frames, where C again is the coding quality control parameter and is between 0 and 1.
The adjustment gain G serves to correct any level mismatch between original and decoded signal that might arise from the noise shaping and de-emphasis. This gain is computed as the ratio of the prediction gain of the short-term analysis and synthesis filter coefficients. The prediction gain of an LPC synthesis filter is the square-root of the output energy when the filter is excited by a unit-energy impulse on the input. An efficient way to compute the prediction gain is by first computing the reflection coefficients from the LPC coefficients through the step-down algorithm, and extracting the prediction gain from the reflection coefficients as
where r_k is the k'th reflection coefficient.
Initial values for the quantization gains are computed as the square-root of the residual energy of the LPC analysis, adjusted by the coding quality control parameter. These quantization gains are later adjusted based on the results of the prediction analysis.
In the prefilter the input signal is filtered using the spectral valley de-emphasis filter coefficients from the noise shaping analysis, see . By applying only the noise shaping analysis filter to the input signal, it provides the input to the noise shaping quantizer.
The prediction analysis is performed in one of two ways depending on how the pitch estimator classified the frame. The processing for voiced and unvoiced speech are described in and , respectively. Inputs to this function include the pre-whitened signal from the pitch estimator, see .
For a frame of voiced speech the pitch pulses will remain dominant in the pre-whitened input signal. Further whitening is desirable as it leads to higher quality at the same available bit-rate. To achieve this, a Long-Term Prediction (LTP) analysis is carried out to estimate the coefficients of a fifth order LTP filter for each of four sub-frames. The LTP coefficients are used to find an LTP residual signal with the simulated output signal as input to obtain better modelling of the output signal. This LTP residual signal is the input to an LPC analysis where the LPCs are estimated using Burgs method, such that the residual energy is minimized. The estimated LPCs are converted to a Line Spectral Frequency (LSF) vector, and quantized as described in . After quantization, the quantized LSF vector is converted to LPC coefficients and hence by using these quantized coefficients the encoder remains fully synchronized with the decoder. The LTP coefficients are quantized using a method described in . The quantized LPC and LTP coefficients are now used to filter the high-pass filtered input signal and measure a residual energy for each of the four subframes.
For a speech signal that has been classified as unvoiced there is no need for LTP filtering as it has already been determined that the pre-whitened input signal is not periodic enough within the allowed pitch period range for an LTP analysis to be worth-while the cost in terms of complexity and rate. Therefore, the pre-whitened input signal is discarded and instead the high-pass filtered input signal is used for LPC analysis using Burgs method. The resulting LPC coefficients are converted to an LSF vector, quantized as described in the following section and transformed back to obtain quantized LPC coefficients. The quantized LPC coefficients are used to filter the high-pass filtered input signal and measure a residual energy for each of the four subframes.
The purpose of quantization in general is to significantly lower the bit rate at the cost of some introduced distortion. A higher rate should always result in lower distortion, and lowering the rate will generally lead to higher distortion. A commonly used but generally sub-optimal approach is to use a quantization method with a constant rate where only the error is minimized when quantizing.Instead, we minimize an objective function that consists of a weighted sum of rate and distortion, and use a codebook with an associated non-uniform rate table. Thus, we take into account that the probability mass function for selecting the codebook entries are by no means guaranteed to be uniform in our scenario. The advantage of this approach is that it ensures that rarely used codebook vector centroids, which are modelling statistical outliers in the training set can be quantized with a low error but with a relatively high cost in terms of a high rate. At the same time this approach also provides the advantage that frequently used centroids are modelled with low error and a relatively low rate. This approach will lead to equal or lower distortion than the fixed rate codebook at any given average rate, provided that the data is similar to the data used for training the codebook.
Instead of minimizing the error in the LSF domain, we map the errors to better approximate spectral distortion by applying an individual weight to each element in the error vector. The weight vectors are calculated for each input vector using the Inverse Harmonic Mean Weighting (IHMW) function proposed by Laroia et al., see .
Consequently, we solve the following minimization problem, i.e.,
where LSF_q is the quantized vector, LSF is the input vector to be quantized, and c is the quantized LSF vector candidate taken from the set C of all possible outcomes of the codebook.
We arrange the codebook in a multiple stage structure to achieve a quantizer that is both memory efficient and highly scalable in terms of computational complexity, see e.g. . In the first stage the input is the LSF vector to be quantized, and in any other stage s > 1, the input is the quantization error from the previous stage, see .
By storing total of M codebook vectors, i.e.,
where M_s is the number of vectors in stage s, we obtain a total of
possible combinations for generating the quantized vector. It is for example possible to represent 2^36 uniquely combined vectors using only 216 vectors in memory, as done in SILK for voiced speech at all sample frequencies above 8 kHz.
This number of possible combinations is far too high for a full search to be carried out for each frame so for all stages but the last, i.e., s smaller than S, only the best min( L, Ms ) centroids are carried over to stage s+1. In each stage the objective function, i.e., the weighted sum of accumulated bit-rate and distortion, is evaluated for each codebook vector entry and the results are sorted. Only the best paths and the corresponding quantization errors are considered in the next stage. In the last stage S the single best path through the multistage codebook is determined. By varying the maximum number of survivors from each stage to the next L, the complexity can be adjusted in real-time at the cost of a potential increase when evaluating the objective function for the resulting quantized vector. This approach scales all the way between the two extremes, L=1 being a greedy search, and the desirable but infeasible full search, L=T/MS. In fact, a performance almost as good as what can be achieved with the infeasible full search can be obtained at a substantially lower complexity by using this approach, see e.g. .
If the input is stable, finding the best candidate will usually result in the quantized vector also being stable, but due to the multi-stage approach it could in theory happen that the best quantization candidate is unstable and because of this there is a need to explicitly ensure that the quantized vectors are stable. Therefore we apply a LSF stabilization method which ensures that the LSF parameters are within valid range, increasingly sorted, and have minimum distances between each other and the border values that have been pre-determined as the 0.01 percentile distance values from a large training set.
The vectors and rate tables for the multi-stage codebook have been trained by minimizing the average of the objective function for LSF vectors from a large training set.
For voiced frames, the prediction analysis described in resulted in four sets (one set per subframe) of five LTP coefficients, plus four weighting matrices. Also, the LTP coefficients for each subframe are quantized using entropy constrained vector quantization. A total of three vector codebooks are available for quantization, with different rate-distortion trade-offs. The three codebooks have 10, 20 and 40 vectors and average rates of about 3, 4, and 5 bits per vector, respectively. Consequently, the first codebook has larger average quantization distortion at a lower rate, whereas the last codebook has smaller average quantization distortion at a higher rate. Given the weighting matrix W_ltp and LTP vector b, the weighted rate-distortion measure for a codebook vector cb_i with rate r_i is give by
where u is a fixed, heuristically-determined parameter balancing the distortion and rate. Which codebook gives the best performance for a given LTP vector depends on the weighting matrix for that LTP vector. For example, for a low valued W_ltp, it is advantageous to use the codebook with 10 vectors as it has a lower average rate. For a large W_ltp, on the other hand, it is often better to use the codebook with 40 vectors, as it is more likely to contain the best codebook vector.
The weighting matrix W_ltp depends mostly on two aspects of the input signal. The first is the periodicity of the signal; the more periodic the larger W_ltp. The second is the change in signal energy in the current subframe, relative to the signal one pitch lag earlier. A decaying energy leads to a larger W_ltp than an increasing energy. Both aspects do not fluctuate very fast which causes the W_ltp matrices for different subframes of one frame often to be similar. As a result, one of the three codebooks typically gives good performance for all subframes. Therefore the codebook search for the subframe LTP vectors is constrained to only allow codebook vectors to be chosen from the same codebook, resulting in a rate reduction.
To find the best codebook, each of the three vector codebooks is used to quantize all subframe LTP vectors and produce a combined weighted rate-distortion measure for each vector codebook and the vector codebook with the lowest combined rate-distortion over all subframes is chosen. The quantized LTP vectors are used in the noise shaping quantizer, and the index of the codebook plus the four indices for the four subframe codebook vectors are passed on to the range encoder.
The noise shaping quantizer independently shapes the signal and coding noise spectra to obtain a perceptually higher quality at the same bitrate.
The prefilter output signal is multiplied with a compensation gain G computed in the noise shaping analysis. Then the output of a synthesis shaping filter is added, and the output of a prediction filter is subtracted to create a residual signal. The residual signal is multiplied by the inverse quantized quantization gain from the noise shaping analysis, and input to a scalar quantizer. The quantization indices of the scalar quantizer represent a signal of pulses that is input to the pyramid range encoder. The scalar quantizer also outputs a quantization signal, which is multiplied by the quantized quantization gain from the noise shaping analysis to create an excitation signal. The output of the prediction filter is added to the excitation signal to form the quantized output signal y(n). The quantized output signal y(n) is input to the synthesis shaping and prediction filters.
Range encoding is a well known method for entropy coding in which a bitstream sequence is continually updated with every new symbol, based on the probability for that symbol. It is similar to arithmetic coding but rather than being restricted to generating binary output symbols, it can generate symbols in any chosen number base. In SILK all side information is range encoded. Each quantized parameter has its own cumulative density function based on histograms for the quantization indices obtained by running a training database.
TBD.
Copy from CELT draft.
Inverse of the post-filter
The MDCT implementation has no special characteristics. The
input is a windowed signal (after pre-emphasis) of 2*N samples and the output is N
frequency-domain samples. A low-overlap window is used to reduce the algorithmic delay.
It is derived from a basic (full overlap) window that is the same as the one used in the Vorbis codec: W(n)=[sin(pi/2*sin(pi/2*(n+.5)/L))]^2. The low-overlap window is created by zero-padding the basic window and inserting ones in the middle, such that the resulting window still satisfies power complementarity. The MDCT is computed in mdct_forward() (mdct.c), which includes the windowing operation and a scaling of 2/N.
The MDCT output is divided into bands that are designed to match the ear's critical
bands for the smallest (2.5ms) frame size. The larger frame sizes use integer
multiplies of the 2.5ms layout. For each band, the encoder
computes the energy that will later be encoded. Each band is then normalized by the
square root of the non-quantized energy, such that each band now forms a unit vector X.
The energy and the normalization are computed by compute_band_energies()
and normalise_bands() (bands.c), respectively.
It is important to quantize the energy with sufficient resolution because
any energy quantization error cannot be compensated for at a later
stage. Regardless of the resolution used for encoding the shape of a band,
it is perceptually important to preserve the energy in each band. CELT uses a
coarse-fine strategy for encoding the energy in the base-2 log domain,
as implemented in quant_bands.c
The coarse quantization of the energy uses a fixed resolution of 6 dB.
To minimize the bitrate, prediction is applied both in time (using the previous frame)
and in frequency (using the previous bands). The prediction using the
previous frame can be disabled, creating an "intra" frame where the energy
is coded without reference to prior frames. An encoder is able to choose the
mode used at will based on both loss robustness and efficiency
considerations.
The 2-D z-transform of
the prediction filter is: A(z_l, z_b)=(1-a*z_l^-1)*(1-z_b^-1)/(1-b*z_b^-1)
where b is the band index and l is the frame index. The prediction coefficients
applied depend on the frame size in use when not using intra energy and a=0 b=4915/32768
when using intra energy.
The time-domain prediction is based on the final fine quantization of the previous
frame, while the frequency domain (within the current frame) prediction is based
on coarse quantization only (because the fine quantization has not been computed
yet). The prediction is clamped internally so that fixed point implementations with
limited dynamic range to not suffer desynchronization. Identical prediction
clamping must be implemented in all encoders and decoders.
We approximate the ideal
probability distribution of the prediction error using a Laplace distribution
with seperate parameters for each frame size in intra and inter-frame modes. The
coarse energy quantization is performed by quant_coarse_energy() and
quant_coarse_energy() (quant_bands.c). The encoding of the Laplace-distributed values is
implemented in ec_laplace_encode() (laplace.c).
After the coarse energy quantization and encoding, the bit allocation is computed
() and the number of bits to use for refining the
energy quantization is determined for each band. Let B_i be the number of fine energy bits
for band i; the refinement is an integer f in the range [0,2^B_i-1]. The mapping between f
and the correction applied to the coarse energy is equal to (f+1/2)/2^B_i - 1/2. Fine
energy quantization is implemented in quant_fine_energy()
(quant_bands.c).
If any bits are unused at the end of the encoding process, these bits are used to
increase the resolution of the fine energy encoding in some bands. Priority is given
to the bands for which the allocation () was rounded
down. At the same level of priority, lower bands are encoded first. Refinement bits
are added until there is no more room for fine energy or until each band
has gained an additional bit of precision or has the maximum fine
energy precision. This is implemented in quant_energy_finalise()
(quant_bands.c).
CELT uses a Pyramid Vector Quantization (PVQ)
codebook for quantizing the details of the spectrum in each band that have not
been predicted by the pitch predictor. The PVQ codebook consists of all sums
of K signed pulses in a vector of N samples, where two pulses at the same position
are required to have the same sign. Thus the codebook includes
all integer codevectors y of N dimensions that satisfy sum(abs(y(j))) = K.
In bands where there are sufficient bits allocated the PVQ is used to encode
the unit vector that results from the normalization in
directly. Given a PVQ codevector y,
the unit vector X is obtained as X = y/||y||, where ||.|| denotes the
L2 norm.
The search for the best codevector y is performed by alg_quant()
(vq.c). There are several possible approaches to the
search with a tradeoff between quality and complexity. The method used in the reference
implementation computes an initial codeword y1 by projecting the residual signal
R = X - p' onto the codebook pyramid of K-1 pulses:
y0 = round_towards_zero( (K-1) * R / sum(abs(R)))
Depending on N, K and the input data, the initial codeword y0 may contain from
0 to K-1 non-zero values. All the remaining pulses, with the exception of the last one,
are found iteratively with a greedy search that minimizes the normalized correlation
between y and R:
J = -R^T*y / ||y||
The search described above is considered to be a good trade-off between quality
and computational cost. However, there are other possible ways to search the PVQ
codebook and the implementors MAY use any other search methods.
The best PVQ codeword is encoded as a uniformly-distributed integer value
by encode_pulses() (cwrs.c).
The codeword is converted from a unique index in the same way as specified in
. The indexing is based on the calculation of V(N,K)
(denoted N(L,K) in ), which is the number of possible
combinations of K pulses in N samples.
When encoding a stereo stream, some parameters are shared across the left and right channels, while others are transmitted separately for each channel, or jointly encoded. Only one copy of the flags for the features, transients and pitch (pitch
period and filter parameters) are transmitted. The coarse and fine energy parameters are transmitted separately for each channel. Both the coarse energy and fine energy (including the remaining fine bits at the end of the stream) have the left and right bands interleaved in the stream, with the left band encoded first.
The main difference between mono and stereo coding is the PVQ coding of the normalized vectors. In stereo mode, a normalized mid-side (M-S) encoding is used. Let L and R be the normalized vector of a certain band for the left and right channels, respectively. The mid and side vectors are computed as M=L+R and S=L-R and no longer have unit norm.
From M and S, an angular parameter theta=2/pi*atan2(||S||, ||M||) is computed. The theta parameter is converted to a Q14 fixed-point parameter itheta, which is quantized on a scale from 0 to 1 with an interval of 2^-qb, where qb is
based the number of bits allocated to the band. From here on, the value of itheta MUST be treated in a bit-exact manner since both the encoder and decoder rely on it to infer the bit allocation.
Let m=M/||M|| and s=S/||S||; m and s are separately encoded with the PVQ encoder described in . The number of bits allocated to m and s depends on the value of itheta.
After all the quantization is completed, the quantized energy is used along with the
quantized normalized band data to resynthesize the MDCT spectrum. The inverse MDCT () and the weighted overlap-add are applied and the signal is stored in the synthesis
buffer.
The encoder MAY omit this step of the processing if it does not need the decoded output.
Each CELT frame can be encoded in a different number of octets, making it possible to vary the bitrate at will. This property can be used to implement source-controlled variable bitrate (VBR). Support for VBR is OPTIONAL for the encoder, but a decoder MUST be prepared to decode a stream that changes its bit-rate dynamically. The method used to vary the bit-rate in VBR mode is left to the implementor, as long as each frame can be decoded by the reference decoder.
It is the intention to allow the greatest possible choice of freedom in
implementing the specification. For this reason, outside of a few exceptions
noted in this section, conformance is defined through the reference
implementation of the decoder provided in Appendix .
Although this document includes an English description of the codec, should
the description contradict the source code of the reference implementation,
the latter shall take precedence.
Compliance with this specification means that a decoder's output MUST be
within the thresholds specified compared to the reference implementation
using the opus_compare.m tool in Appendix .
The codec needs to take appropriate security considerations
into account, as outlined in and .
It is extremely important for the decoder to be robust against malicious
payloads. Malicious payloads must not cause the decoder to overrun its
allocated memory or to take much more resources to decode. Although problems
in encoders are typically rarer, the same applies to the encoder. Malicious
audio stream must not cause the encoder to misbehave because this would
allow an attacker to attack transcoding gateways.
The reference implementation contains no known buffer overflow or cases where
a specially crafter packet or audio segment could cause a significant increase
in CPU load. However, on certain CPU architectures where denormalized
floating-point operations are much slower it is possible for some audio content
(e.g. silence or near-silence) to cause such an increase
in CPU load. For such architectures, it is RECOMMENDED to add very small
floating-point offsets to prevent significant numbers of denormalized
operations or to configure the hardware to zeroize denormal numbers.
No such issue exists for the fixed-point reference implementation.
This document has no actions for IANA.
Thanks to all other developers, including Raymond Chen, Soeren Skak Jensen, Gregory Maxwell,
Christopher Montgomery, Karsten Vandborg Soerensen, and Timothy Terriberry.
SILK Speech Codec
Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantization
Evaluation of Split and Multistage Techniques in LSF QuantizationEfficient Search and Design Procedures for Robust Multi-Stage VQ of LPC Parameters for 4 kb/s Speech CodingConstrained-Energy Lapped Transform (CELT) CodecInternet Denial-of-Service ConsiderationsIABThis document provides an overview of possible avenues for denial-of-service (DoS) attack on Internet systems. The aim is to encourage protocol designers and network engineers towards designs that are more robust. We discuss partial solutions that reduce the effectiveness of attacks, and how some solutions might inadvertently open up alternative vulnerabilities. This memo provides information for the Internet community.Guidelines for Writing RFC Text on Security ConsiderationsAll RFCs are required to have a Security Considerations section. Historically, such sections have been relatively weak. This document provides guidelines to RFC authors on how to write a good Security Considerations section. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements.Range encoding: An algorithm for removing redundancy from a digitised messageSource coding algorithms for fast data compressionA Pyramid Vector QuantizerThis appendix contains the complete source code for the
reference implementation of the Opus codec written in C. This
implementation can be compiled for
either floating-point or fixed-point architectures.
The implementation can be compiled with either a C89 or a C99
compiler. It is reasonably optimized for most platforms such that
only architecture-specific optimizations are likely to be useful.
The FFT used is a slightly modified version of the KISS-FFT package,
but it is easy to substitute any other FFT library.
The complete source code can be extracted from this draft, by running the
following command line:
opus_source.tar.gz
]]>
tar xzvf opus_source.tar.gz
cd opus_sourcemake