IPR2017-01075
`Saint Lawrence Communications
`Exhibit 2003
`
`PAGE 249
`
`Lossless Wideband Speech Coding
`
`C. H. Ritz, J. Parsons
`TITR, University of Wollongong, NSW, Australia
`chritz@elec.uow.edu.au
`
`Abstract
`
`This paper investigates lossless coding of wideband speech by adding a
`lossless enhancement layer to the lossy baselayer produced by a standardised
`wideband speech coder. Both the ITU-T G.722 and G.722.2 speech coders
`are examined. Entropy results show that potential compression rates are
`dependent on the type and bit rate of the baselayer coder as well as the
`symbol size used by the lossless coder. Higher compression rates were
`obtained by adding a decorrelation stage prior to lossless encoding. The
`resulting lossless speech coder operates at a bit rate that is approximately
`58% of the bit rate of original digitised wideband speech signal.
`
`1. Introduction
`
`Wideband speech refers to speech sampled at 16 kHz
`and offers superior quality to narrowband speech
`traditionally
`used
`in
`telephony
`applications
`(Mermelstein, 1988). Existing research into wideband
`speech coding has mainly focused on lossy coding and
`has resulted in the standardisation of numerous speech
`coding algorithms. Lossy speech coders aim
`to
`reproduce a “perceptually lossless” version of the
`speech signal at bit rates much lower than the original
`digitised version. In contrast, lossless speech coding
`aims to allow full reproduction of the original digitised
`speech signal and is the focus of this work. Lossless
`speech coding finds applications where only perfect
`reconstruction is tolerated. Examples include speech
`storage for future editing in the recording or movie
`industries and archiving of
`legal proceedings or
`historical events.
` While there have been numerous proposals for
`lossless coding of audio signals (Hans and Schafer,
`2001), there is only limited research into lossless coding
`specifically for speech signals. Many of these existing
`approaches rely on firstly removing correlation from the
`speech signal (using for example Linear Prediction
`(LP)), before encoding the resulting signal with a
`lossless encoder. Lossless speech coders combining LP
`with dynamic Huffman coding (Ramsey and Gribble,
`1987; Garafolo, Robinson and Fiscus, 1994), arithmetic
`coding (Stearns, 1995), and Golomb-Rice encoding
`(Giurcaneanu, Tabus and Astola, 2000) have been
`proposed. The lowest compression rate (defined as the
`ratio of lossless bit rate to original bit rate) achieved by
`these proposals was 43-45% when operating of 16 kHz
`sampled speech (Garafolo, Robinson and Fiscus, 1994;
`Giurcaneanu, Tabus and Astola, 2000)). Results for
`lossless audio coders applied to speech signals have
`achieved compression rates down to 33% (Li, 2003),
`
`IPR2017-01075
`Saint Lawrence Communications
`Exhibit 2003
`
`although these results were for speech sampled at 44.1
`kHz; hence overall bit rates are higher than those
`reported for wideband speech.
` The focus of this research is to investigate lossless
`speech coding by providing a lossless enhancement
`layer for an existing standardised wideband speech
`coder. The motivation for this approach is to allow for
`both a lossy and lossless version to be available in the
`one scheme. The
`two wideband speech coders
`investigated in this paper are; the ITU-T G.722 speech
`coder (Mermelstein, 88) and the Adaptive Multi Rate -
`Wideband (AMR-WB) speech coder (Bessette, et. al,
`2002), standardised as ITU-T G.722.2.
`
`Section 2 of this paper describes the overall lossless
`coding scheme proposed. Section 3 examines the
`residual characteristics of the chosen wideband speech
`coders in the context of lossless coding. Section 4
`presents results of methods for reducing the bit rate
`required for lossless residual compression while Section
`5 presents results for a practical lossless wideband
`speech coder and compares results with existing lossless
`coders. Conclusions are presented in Section 6.
`
`2. Lossless coder structure
`
`The structure of the coder proposed here is illustrated in
`Figure 1. In Figure 1, the speech signal is first encoded
`with the lossy speech encoder producing a lossy
`bitstream denoted bly. The difference between the
`original and reconstructed speech signal from the lossy
`coder, referred to as the residual, is then encoded using
`the lossless encoding stage which produces a second
`bitstream denoted bls. In the decoder, both the lossy
`speech and residual bitstreams are decoded using the
`lossy
`and
`lossless decoders
`respectively. The
`synthesized speech signal is added to the recovered
`residual signal to form the original speech signal.
`
`Proceedings of the 10th Australian International Conference on Speech Science & Technology
`Macquarie University, Sydney, December 8 to 10, 2004. Copyright, Australian Speech Science & Technology Association Inc.
`
`Accepted after full paper review
`
`IPR2016-00704
`SAINT LAWRENCE COMMUNICATIONS LLC
`Exhibit 2003
`
`

`

`s(n)
`
`Lossy
`Encoder
`
`Lossy
`Decoder
`
`bly
`
`+
`
`bls
`
`Lossless
`Encoder
`
`+
`
`s(n)
`
`(a)
`
`bly
`
`bls
`
`(b)
`
`Lossy
`Decoder
`
`Lossless
`Decoder
`
`Figure 1. Lossless coding scheme. (a) Encoder,
`(b) Decoder
`
`
`
`
`
`3. Residual characteristics of G.722 and
`G.722.2
`
`Both wideband speech coders selected can operate at
`multiple bit rates: 48 kbps, 56 kbps and 64 kbps for
`G.722; and nine different bit rates ranging from 6.6
`kbps to 23.85 kbps for G.722.2. To determine which
`coder would result in the lowest potential lossless rate,
`the entropy was measured for residuals obtained using
`different bit rates and coders for the baseline lossy
`coder. All speech used in this work was taken from the
`ANDOSL database (ANDOSL, 1998), bandlimited to
`the range 50 Hz to 7 kHz and resampled to 16 kHz. The
`entropy was measured in bits per sample using the
`equation:
`
`N
`
`H
`
`= ∑
`=
`1
`
`p
`i
`
`log
`
`2
`
`p
`i
`
`
`
`
`
`
`
`
`
`
`
`(1)
`
`
`
`i
`
`
`
`In expression (1), N is the alphabet size and pi is the
`probability of each symbol. Note that the residual
`signals were assumed to be generated by an independent
`and
`identically distributed
`(iid) source and so
`probabilities were approximated as the frequency of
`occurrence of each symbol. Hence, this is an upper limit
`for the entropy.
`
`Initial entropy calculations assumed 16 bit symbol
`sizes corresponding to the 16 bit residual samples. An
`investigation was performed into the benefits of using
`smaller or larger symbol sizes. Specifically, 16 bit
`residual samples were broken up into 8 bit symbols and
`also combined to form 32 bit symbols. Tables 1 and 2
`present results for entropy of the residual signals in bits
`per 16 kHz sample using different symbols sizes and
`averaged over four speech files for the G.722 and
`G.722.2 coders, respectively.
`
`
`
`
`Entropies (bits/sample)
`
`PAGE 250
`
`48
`
`56
`
`64
`
`Symbol
`Size (bits)
`8
`16
`
`8.7155
`9.1
`9.5
`6.44575
`6.8
`7.3
`5.646
`5.9
`6.2
`32
`Table 1. Entropy results for variation of symbol
`sizes for different G.722 bit rates (kbps).
`
`Entropies (bits/sample)
`
`6.6
`
`14.25
`
`19.85
`
`23.85
`
`
`Symbol
`Size (bits)
`12.1
`12.2
`12.2
`12.3
`8
`10.3
`10.3
`10.3
`10.5
`16
`7.0
`6.9
`6.9
`7.0
`32
`Table 2. Entropy results for variation of symbol
`sizes for different G.722.2 bit rates (kbps).
`
`
`Table 1 shows that the entropy in bits per sample
`decreases as the bit rate of the lossy coder increases for
`all symbol sizes. This result is reasonable as G.722 is an
`ADPCM coding scheme and hence residuals will have
`decreased dynamic range and hence entropy as the bit
`rate and hence coding accuracy increases. Table 1 also
`shows that the entropy decreases as the symbol size
`increases. The decreased entropy for 16 bit symbols as
`opposed to 8 bit symbols indicates correlation may be
`present between the Most Significant Bytes (MSBs) and
`Least Significan Bytes (LSBs) of each sample. The
`further decrease in entropy for 32 bit symbols could
`indicate correlation between adjacent residual samples.
`
`In contrast to Table 1, Table 2 shows the entropy in
`bits per sample for a given symbols size is relatively
`similar for all bit rates of the lossy coder. This could be
`explained by the fact that G.722.2 uses a Linear
`Prediction (LP) model and an analysis-by-synthesis
`coding scheme. Hence, the residual signal will follow a
`similar shape to the original speech signal resulting in a
`similar dynamic range regardless of bit rate. However,
`similar to the results of Table 1, Table 2 shows that the
`entropy decreases as the symbol size increases. This
`result could be explained using similar reasoning as
`used to explain the corresponding results of Table 1.
` To compare total overall lossless bit rates when
`adding the additional lossy base layer, Figure 2 plots
`potential bit rates for each coder and for each symbol
`size. Here, potential bit rates are estimated as the sum of
`the bit rates of the lossy and lossless stages of Figure1.
`Figure 2 shows that the lowest bit rate results when
`using 32 bit symbol sizes and the G.722.2 coder
`operating at 6.6 kbps. However if 16 or 8 bit symbol
`sizes are selected then the G.722 coder operating at 48
`kbps gives the best performance. For both coders tested,
`the highest lossless compression rate is achieved when
`operating the respective coder at the lowest bit rate.
`
`
`Proceedings of the 10th Australian International Conference on Speech Science & Technology
`Macquarie University, Sydney, December 8 to 10, 2004. Copyright, Australian Speech Science & Technology Association Inc.
`
`Accepted after full paper review
`
`

`

`PAGE 251
`
`r(n)
`
`Decorrelation
`
`Symbol
`splitter
`
`Lossless
`Decoder
`
`Lossless
`Decoder
`
`(a)
`
`bls1
`
`bls2
`
`(b)
`
`bls1
`
`bls2
`
`Lossless
`Encoder
`
`Lossless
`Encoder
`
`Symbol
`combiner
`
`Recorrelation
`
`r(n)
`
`
` Figure 3. Enhanced lossless coder using symbol
`
`splitting and decorrelation. (a) Encoder (b) Decoder.
`
`8 bit symbols
`16 bit symbols
`32 bit symbols
`
`220
`
`200
`
`180
`
`160
`
`140
`
`120
`
`100
`
`Lossless bit rate
`
`48
`
`56
`
`64
`
`6.6
`
`14.25
`
`19.85
`
`23.85
`
`Lossy baselayer Bit Rate
`
`
`
`Figure 2. Potential bitrates for different symbols
`sizes and lossy coders. 48-64 kbps: G.722; 6.6-
`23.85 kbps: G.722.2.
`
`4. Symbol Splitting and Correlation Reduction
`
`
`
`Results of Section 3 indicate that using larger symbol
`sizes should lead to a coding gain. However, increasing
`the symbol size leads to an increase in alphabet size. For
`a practical lossless coder, e.g. a Huffman coder, this
`requires formation and transmission of a larger Huffman
`table, hence increasing processing time and lossless rate.
`Here, a compromise using symbol splitting
`is
`investigated.
` Results from Section 3 also showed that removing
`correlation should also lead to a coding gain. Here,
`decorrelation of the residual is investigated using first
`order linear prediction via Differential Pulse Code
`Modulation (DPCM).
`
`4.3. Combined approach
`
`To take advantage of both symbol splitting decorrelation
`using DPCM, a combined approach was also
`investigated. The overall structure of the lossless coder
`for the residual using this combined approach is shown
`in Figure 3.
`
`In Figure 3, the residual, r(n) is passed through a
`decorrelation stage, in this case DPCM. The resulting
`decorrelated signal samples are then split into two new 8
`bit
`symbols using
`the
`first
`splitting
`technique
`investigated in Section 4.1. Each of these symbols are
`encoded with separate lossless encoders to produce two
`lossless bit streams bls1 and bls2.
`
`4.1. Symbol Splitting
`
`4.4. Results
`
`The technique of symbol splitting involves separating
`larger symbols into smaller symbols and using a
`separate lossless coder for each new symbol.
` Two methods of splitting the 16 bit samples and
`combining to form new symbols were investigated. The
`first method split the 16 bit samples into two 8 bit
`symbols formed from the LSB and MSB. This aimed to
`exploit the observation that the residual samples were
`mostly of small magnitude, hence the LSBs and MSBs
`should have different lossless coding requirements. The
`second method combined the adjacent 16 bit samples to
`form two new 16 bit symbols. Each new symbol was
`formed by combining the adjacent MSBs and LSBs,
`respectively in an attempt to exploit sample-to-sample
`correlation.
`
`4.2. Decorrelation via DPCM
`
`In this approach, the residual signal is predicted using
`DPCM. A new residual is obtained as the difference
`between the DPCM synthesised residual and the original
`residual; this new signal is then losslessly coded. This
`approach aims to exploit correlation between adjacent
`signal samples as indicated from results of Section 3.
`
`Figure 4 shows the resulting overall minimum bit rates
`when measuring the entropy of these new methods for
`the same speech files as presented in Figure 2. For
`comparison purposes, results are re-plotted when using
`the original 16 bit and 32 bit symbol sizes.
`
`Figure 4 shows that the split method with 16 bit
`symbols results in a lower bit rate than both the split 8
`bit symbols and the original 16 bit symbols. Comparing
`with Figure 2, results for split 8 bit symbols are also
`superior to those for the original 8 bit symbols.
`
`Figure 4 also shows that adding the DPCM stage
`results in a significant reduction in the overall lossless
`bit rate for the G.722.2 coder but increases the bit rate
`for the G.722 coder. Results for DPCM and 8-bit split
`symbols are similar to results obtained without DPCM
`but using 16 bit split symbols.
` While results for 32 bit symbols are still superior, to
`avoid the practical problems of creation and storage of
`large Huffman tables, the method utilising a DPCM
`stage and 8 bit symbol sizes was chosen for the
`remainder of the work presented here. Based on the
`results of Figure 4, the G.722.2 coder operating at 6.6
`kbps was chosen as the baseline lossy coder.
`
`Proceedings of the 10th Australian International Conference on Speech Science & Technology
`Macquarie University, Sydney, December 8 to 10, 2004. Copyright, Australian Speech Science & Technology Association Inc.
`
`Accepted after full paper review
`
`

`

`Powered by TCPDF (www.tcpdf.org)
`
`PAGE 252
`
`The results of Table 3 also show that the ADPCM
`technique is able to reduce the bit rate by approximately
`10 kbps below the estimated entropy however Monkey’s
`audio results in the lowest overall bit rate. The superior
`performance could be explained by
`the more
`sophisticated decorrelation stage of Monkey’s audio
`compared with the ADPCM technique.
`
`6. Conclusions
`
`This paper has described lossless coding of speech using
`a lossy coder as the baselayer and Huffman coding
`applied to a decorrelated residual signal. Results show
`that approximately 58% compression of the original
`speech signal could be achieved. This result is still
`inferior to the rate achievable by a state of the art
`lossless audio coder.
`
`Future work should focus on two aspects. Firstly, the
`implementation of a lossless coder capable of encoding
`32 bit symbols and avoiding the construction of large
`symbol dictionaries, using, for example, adaptive
`Huffman coding
`(Sayood, 2000). Secondly,
`the
`investigation of more sophisticated techniques for
`removing correlation from the residual signal such as
`non-linear predictive techniques as used in Monkey’s
`audio. In particular, a more thorough investigation into
`the most appropriate correlation reduction technique for
`residual signals obtained from lossy speech coders
`should be performed.
`
`7. References
`
`ANDOSL: Australian National Database of Spoken Language
`CD ROM, 1998.
`Ashland, M. T., “Monkeys Audio, A fast and powerful
`lossless audio compressor”, www.monkeysaudio.com,
`Version 3.97, 2002.
`Bessette, B., et. al, “The adaptive multirate wideband speech
`codec (AMR-WB)”, IEEE Trans. Speech and Audio Proc.,
`Vol.: 10, Issue: 8 , pp. 620 -636, Nov 2002.
`Garafolo, J. S., Robinson, T. and Fiscus, J. G., “The
`development of file formats for very large speech copora:
`SPHERE and SHORTEN”, Proc. ICASSP’94, Vol. I, pp.
`113-116, 1994.
`Giurcaneanu, C. D., Tabus , I. and Astola, J., “Adaptive
`context-based sequential prediction for lossless audio
`compression”, Sig. Proc., Vol. 80, pp. 2283-2294, 2000.
`Hans, M. and Schafer , R. W., “Lossless compression of
`digital audio”, IEEE SP Magazine, pp. 21-32, Jul., 2001.
`Li, J., “A progressive to lossless audio coder (PLEAC) with
`reversible modulated
`lapped
`transform”,
`Proc.
`ICASSP2003, Vol. V, pp. 413-416.
`Mermelstein, P., ‘G.722, A New CCITT Coding Standard for
`Digital Transmission of Wideband Audio Signals’, IEEE
`Comm. Magazine, Vol. 26, No. 1, pp. 8-15, Jan. 1988.
`Ramsey, L. T. and Gribble, D., “Information-theoretic
`compressibility of speech data”, Proc. ICASSP '87, Vol.
`12, p. 17-20, Apr 1987.
`Sayood, K., “Introduction to Data Compression”, 2nd Ed.,
`Academic Press, 2000.
`Stearns, S. D., ‘Arithmetic Coding in Lossless Waveform
`Compression’, IEEE Trans. on Signal Processing, Vol. 3,
`No. 8, pp. 1874-1879, August 1995.
`
`8-bit split symbols
`16-bit split symbols
`original 32 bit symbols
`original 16 bit symbols
`DPCM residual
`
`200
`
`190
`
`180
`
`170
`
`160
`
`150
`
`140
`
`130
`
`120
`
`110
`
`100
`
`Overall Lossless Rate (kbps)
`
`48
`
`56
`
`64
`
`6.6
`
`14.25
`
`19.85
`
`23.85
`
`Lossy baseline Bit rate (kbps)
`
`
`
`Figure 4. Potential bit rates resulting from using
`different symbol sizes and decorrelation. 48-64
`kbps: G.722; 6.6-23.85 kbps: G.722.2.
`
`5. A practical lossless speech coder
`
`Three well known lossless coding methods for the
`residual were investigated: Huffman coding, Arithmetic
`coding and Lempel-Ziv (LZ) coding (Sayood, 2000).
`For the LZ coding method both LZ77 and LZ78 was
`utilised. Results were obtained for 14 speech files
`consisting of male and female sentences encoded using
`the combined approach described in Section 4.3 and the
`6.6 kbps G.722.2 coder.
` A further technique was also investigated whereby
`the DPCM stage described in Section 4.2 was replaced
`by an Adaptive DPCM (ADPCM) stage similar to that
`utilised in the G.722 speech encoder. To compare with
`an existing state of the art lossless audio coder, results
`were also obtained when coding the residual signals
`using Monkey’s audio (Ashland, 2002). Monkey’s audio
`was chosen as it has been shown to provide superior
`performance over many other lossless audio coders (Li,
`2003). Entropy based compression rates for overall
`compression rates using each lossless coding technique
`are shown in Table 3.
`
`Lossless Bit Rate (kbps)
`Coding Method
`160.9
`Huffman
`159.3
`Arithmetic
`173.3
`LZ77
`170.6
`LZ78
`158.7
`Entropy
`Huffman with ADPCM 149
`Monkey’s audio
`123
`Table 3. Resulting bit rates for a practical
`lossless coder.
`
`The results of Table 3 show that both Arithmetic and
`Huffman coding achieve similar results that are superior
`to the LZ results. The inferior results of LZ techniques
`agree with other authors (Giurcaneanu, Tabus and
`Astola, 2000; Garafolo, Robinson and Fiscus, 1994).
`
`Proceedings of the 10th Australian International Conference on Speech Science & Technology
`Macquarie University, Sydney, December 8 to 10, 2004. Copyright, Australian Speech Science & Technology Association Inc.
`
`Accepted after full paper review
`
`