Super-Wideband Fine Spectrum Quantization for Low-rate High-Quality MDCT Coding Mode of The 3GPP EVS Codec

Transcription

1 Super-Wideband Fine Spectrum Quantization for Low-rate High-Quality DCT Coding ode of The 3GPP EVS Codec Presented by Srikanth Nagisetty, Hiroyuki Ehara 15 th Dec 2015

2 Topics of this Presentation Background EVS Encoder Overview LR-DCT Coder Envelope Coding Spectrum Coding Bit allocation Gap Filling Evaluation Results Summary 2

3 Background Challenge: Encoding SWB band signal at low delay & low bit-rate ixed contents and music sampled at 32kHz Frame length : 20 ms Too few bits for quantizing SWB spectrum (quantizing 560 bins using around 256 bits) Highly efficient quantization algorithm is needed 3

4 EVS Encoder Overview Speech Signals ixed & usic Signals The Low Rate High Quality (LR-HQ) DCT coding is one of the mode in the EVS DCT coder. 4

5 LR-HQ DCT SWB Encoder (1/2) Quantized LF information for gap filling Signal Classification Energy Coding Spectrum Coding Transient Normal Harmonic Huffman Coding Advanced bit allocation 5

6 I Scaled Energies E ( b) round( q ( b) ), int LR-HQ DCT SWB Coder b 0,..., N bands -Envelope coding 1 Selection based on 1. Range of indices 2. Bits consumption E ( b) log 2 k kend ( b) 1 X k k start ( b) ( k) 2 Epsilon, Band Energies b 0,..., N bands 1 I I (0) I ( b) I I (0) round( q ( b) I ref int ( b 1), ) b 1,..., N Differential Indices bands 1 6

7 Energy Spectral Peak Tracking Peak Average Energy Average-Energy HF Bands n HF bands for Quantization Non-Zero Non-Zero Bit allocation Reduces the width for the current frame - Bit savings - Inter-frame Continuity Spectral Peak Tracking Bit allocation Temporal discontinuity -If the current frame HF spectrum peak is quantized while the previous frame is not quantized -Perceivable degradation 7

8 LR-HQ DCT Coder -Bit allocation (1/2) Dynamic Bit allocation Bands(1) Bits(1) Bits(1)>T1 Bands(2) : Bands(n-1) Dynamic Bit allocation Bits(2) : Bits(n-1) Bits(2)>T2 : Bits(n-1)>Tn-1 Combination Logic Bit allocation Result Bands(n) Bits(n) Bits(n)>Tn Iterative Bands with bits < Threshold identified and relocates bits to other bands 8

9 Bands(1) Bands(2) : Bands(n-1) Bands(n) LR-HQ DCT Coder -Bit allocation (2/2) Adaptive bit allocation based on Grouping Adaptive Grouping Identifies dominant bands (local max) and its valley to form groups Group(1) Group(2) : Group(k) Group energies Identify Dominant Group Bit allocation to Group Dynamic Bit allocation to bands in a Group Bit allocation Iterative: Identify Next Dominant Group Group energy and Variance Bit allocation Results Bits allocated 1) Adaptively grouping the bands and 2) By exploiting the relationship between the groups. This approach is more suitable for tonal (Harmonic) like signals as the energy of the bands is mainly concentrated at discrete tones. 9

10 LR-HQ DCT Coder -Gap Filling khz TCQ/USQ is performed on Non Zero bands Noise-filling is performed on 0-bit-allocated LF bands Spectral Gaps Zero-bit bands cause spectral gaps, which lead to audible artifacts if left alone. Gap filling technique is used. 10

11 LR-HQ DCT Coder -Gap Filling Normal ode khz TCQ/USQ is performed on Non Zero bands Noise-filling is performed on 0-bit-allocated LF bands Best match spectrum filled in the High Frequency Normalized LF 11

12 LR-HQ DCT Coder -Gap Filling Harmonic ode khz TCQ/USQ is performed on Non Zero bands Normalized LF 1.Harmonic relation not maintained b/w reproduced HF and input HF 2.Roughness perceived during the boundary regions Noise-filling is performed on 0-bit-allocated LF bands Gap Filled Spectrum Best match spectrum filled in the High Frequency irroring 12

13 f(n) Last spectral peak in LF LR-HQ DCT Coder -Spectral Correction Roughness perceived when 1 st spectral peak in HF 2 tones fall in 3 rd 30 spectral Hz and amplitudes peak are in rapidly HF changing f(n) Synthesized LF spectrum Last spectral peak in LF 1 st spectral peak in HF EstHarmonic Est Harmonic Replicated HF spectrum 2 nd spectral peak in HF Est Harmonic 3 rd spectral peak in HF freq Spectral correction using Harmonic transposition. This method retains both the fine structure of the spectrum and the harmonic relationships between the low frequency tones and the replicated high frequency tones 13 freq

14 LR-HQ DCT Coder - Listening Test Setup Degradation Category Rating (DCR) methodology (ITU-T P.800) : 24 ixed and usic samples recorded in Japanese language 16 Japanese naïve listeners Codecs EVS SWB LR-HQ mode Reference Codec: AR-WB+ 14

15 LR-HQ DCT Coder - Evaluation Results LR-HQ SWB Performance is equal or greater than AR-WB+ whose algorithmic delay is longer than twice of EVS (32ms) 15

16 Summary For encoding the SWB spectral coefficients at low bit budget Spectral band energies are quantized using an efficient Huffman coding methods Advanced bit allocation methods are used for efficient representation of spectrum. Spectral holes in the full spectrum coding is filled using gap filling techniques Gap-filling techniques are improved by introducing a fine spectrum normalization and adaptive sparse BWE coding Conclusion: EVS LR-HQ SWB coder meets the performance requirements and is adopted as a part of multi-mode DCT coding in the EVS codec. 16

17 References 1. 3GPP TR Study of Use Cases and requirements for enhanced voice codecs for the EPS, v , ar GPP TS andatory speech codec speech processing functions; Adaptive ulti-rate (AR) speech codec; General description Speech, v , Sep GPP TS Speech codec speech processing functions; Adaptive ulti-rate-wide-band (AR-WB) speech codec; General description, v , Oct GPP TS Codec for Enhanced Voice Services (EVS); Detailed Algorithmic Description, v ), Sep S. Bruhn, et al., Standardization of the new EVS Codec, IEEE ICASSP, April S. Nagisetty, et al., Low Bitrate High Quality DCT Audio coding of the 3GPP EVS, IEEE ICASSP ITU-T G.719, Low complexity, full band audio coding for high quality, conversational applications, ITU-T Recommendation G.719, June T. R. Fischer, H. Sung, J. Zhan, E. Oh, High-quality audio transform coded excitation using trellis codes, ICASSP, 2008, pp ITU-T G.718 Annex B, Superwideband scalable extension for ITU-T G.718, ITU-T Recommendation G.718 Amendment 2, arch, L. Laaksonen et al., Superwideband extension of G.718 and G speech codecs, Interspeech Oshikiri, H. Ehara, K. Yoshida, Efficient spectrum coding for super-wideband speech and its application to 7/10/15 khz bandwidth scalable coders, IEEE ICASSP, 2004, pp.i H. Fastl, E. Zwicker, Psychoacoustics: Facts and odels. Springer series in information sciences. Springer, 3 rd edition, Audio codec processing functions; Extended Adaptive ulti-rate Wide-band (AR-WB+) codec, 3GPP TS GPP, T-doc S , EVS performance requirements, April ITU-T P.800, ethods for Subjective Determination of Transmission Quality. International Telecommunication Union (ITU), Series P., August GPP, Tdoc S , Report of the Global Analysis Lab for the EVS Selection Phase, Aug

18 APPENDIX 18

19 LR-HQ DCT SWB Encoder (2/2) Energy Frequency khz Spectral Coefficients are grouped Spectral bands E(1) E(2).. E(b-1) E(b) Spectral band energies 19

20 8 Symbol Huffman table LR-HQ DCT Coder -Envelope coding (2/2) I(0) : I(b) 32 Symbol Huffman tables Large Symbol Huffman > [-64 63] Range Small Symbol Huffman < [-64 63] Pulse ode Scale ode Context Based Huffman Resized Huffman If I(b) exceeds [-4 3] 1. Position 2. Amplitude of coded directly I(b) I(b-1) determines best Huffman table for encoding the current band I(b) I(b) Span is narrowed down to fewer code words (21 symbols) 20

21 LR-HQ DCT Coder -Envelope coding (3/3) Small Symbol coding method Context based Huffman: table for encoding the current band Resized Huffman : I(b) I(b-1) determines best Huffman I(b) narrowed to a smaller range for using Huffman table with fewer symbols (21 symbols). ΔI (b) ΔI ΔI (b) (b) min( ΔI max( ΔI (b-1 ) (b-1 ) -T,3), 1 ΔI, 3), ΔI (b-1 ) T (b-1 ) - T T ΔI (b) is the the new differenti al index for band b T 15 thr, T 1 15 thr 21

22 LR-HQ DCT Coder -Quantization (TCQ and USQ) agnitude Quantized by USQ & TCQ Position, Number and sign coded by Arithmetic coding 22

23 LR-HQ DCT Coder -Sparse Band Search Reduces computational load Best match spectrum filled in the High Frequency Region Best match index is identified by finding the k' which maximizes the correlation measure, S(k') according to S( k' ) corr( k' ) corr( k') Ene( k') Ncnt 1 k0 Ncnt 1 ~ X k0 2 Ene( k' ), k' 0,..., Nlag 1 ~ X ( Idx[ k]) X ( k lag[ k'] Idx[ k]) ( k lag[ k'] Idx[ k]) 2 Gap filling for Normal mode 23