1. MEMORY-LESS GAIN QUANTIZATION IN THE EVS CODEC Vladimir Malenovsky Milan Jelinek University of Sherbrooke/VoiceAge Corp. CANADA

2. Motivation and Design Constraints eliminate the propagation of energy mismatch in the decoder due to incorrect predicted gain of the fixed excitation signal consume the same or lower amount of bits as the AMR-WB gain quantizer complexity not higher than few WMOPS memory consumption not higher than 5k words subjective quality same or better than that of AMR-WB gain quantizer implementation in the native modes of the EVS codec

3. Gain quantization in the AMR-WB codec both gains are quantized jointly by a 2-entry VQ the gain of the fixed excitation is predicted from previous frames in the log. domain 6-7 bits per subframe

4. Gain prediction in the AMR-WB codec

5. Error propagation the FEC mechanism in the decoder estimates the evolution of the gain during frame losses in stable frames (VOICED/UNVOICED), the gain is almost unchanged in TRANSITION frames, the gain is attenuated towards zero in the recovery stage, the predicted gain could be wrong and consequently the decoded gain could be wrong -> energy undershoots/overshoots

6. Intra-frame gain prediction legend G c0 – predicted gain, G i – innovation gain, G c – unquantized gain, Ĝ c – quantized gain, Γ – gain correction factor a / b – vector/matrix of prediction coefficients p i / P i – predictor vector/matrix 1 st subframe the only predictor is the coder type (GENERIC, VOICED, UNVOICED,...) 2 nd, 3 rd,... subframes quantized gains ĝ p and Ĝ c from previous subframe(s) are taken as predictors in the next subframe(s) sub-frame gains are quantized with 6 bits

7. Estimation of target signal energy the target signal energy subtract an estimate of the adaptive excitation energy and calculate the predicted gain G tp is quantized with 5 bits and sub-frame gains are quantized with 6 bits

8. Evaluation and testing clean speech -26 dBov, 12.8 kHz ACELP synthesis bitratemethodAMR-WBEVS IEVS IIEVS I – AMR-WB EVS II – AMR-WB 6.6 SNR7.217.988.050.770.84 SSNR6.486.937.030.450.55 8.85 SNR9.5310.1210.150.590.62 SSNR8.418.788.870.370.46 12.65 SNR12.3712.5312.60.160.23 SSNR11.0210.9311.08-0.090.06 18.05 SNR15.0314.9515.12-0.080.09 SSNR13.4813.1813.44-0.30-0.04 23.05 SNR16.7216.4916.76-0.230.04 SSNR15.0814.6515.03-0.43-0.05

9. Evaluation and testing MUSHRA test (6.6 kbps) clean speech at -26, -16 and -36 dBov 16 randomly selected items EVS I and EVS II >> AMR-WB EVS I ~= EVS II two blind AB tests (6.6 and 7.5 kbps) clean speech at -26, -16 and -36 dBov expert listeners EVS I > EVS II

10. Conclusion two memory-less gain quantization schemes were developed and implemented in the EVS codec intra-frame gain prediction (5.9 VBR, 7.2 and 8.0 kbps) estimate of the target signal energy (9.6 – 32 kbps) the problem of energy mismatch propagation in the decoder due to incorrect predicted gain of the fixed excitation signal has been eliminated the new quantizers are bitrate-scalable both outperform the AMR-WB gain quantizer at bitrates below 12.65 kbps and are on par with AMR-WB at bitrates above 12.65 kbps the first proposed scheme has been preferred by expert listeners at very low bitrates complexity and memory consumption is similar to the gain quantizer of AMR- WB (it’s slightly higher in case of intra-frame gain prediction)

11. DONE