wavenet-coding 0.2.0__tar.gz → 0.2.1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (28) hide show
  1. {wavenet_coding-0.2.0/wavenet_coding.egg-info → wavenet_coding-0.2.1}/PKG-INFO +41 -42
  2. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/README.md +40 -41
  3. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/scripts/evaluate.py +82 -19
  4. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/scripts/inference.py +95 -4
  5. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/scripts/prepare_data.py +20 -2
  6. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/scripts/train.py +144 -72
  7. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/setup.py +1 -1
  8. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/__init__.py +8 -1
  9. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/codec2.py +80 -172
  10. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/evaluation.py +30 -26
  11. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/inference.py +90 -81
  12. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/model.py +295 -53
  13. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1/wavenet_coding.egg-info}/PKG-INFO +41 -42
  14. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/LICENSE +0 -0
  15. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/MANIFEST.in +0 -0
  16. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/pyproject.toml +0 -0
  17. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/requirements.txt +0 -0
  18. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/scripts/export_model.py +0 -0
  19. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/setup.cfg +0 -0
  20. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/configs.py +0 -0
  21. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/dataset.py +0 -0
  22. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/entropy_coding.py +0 -0
  23. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/export.py +0 -0
  24. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding/loss.py +0 -0
  25. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding.egg-info/SOURCES.txt +0 -0
  26. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding.egg-info/dependency_links.txt +0 -0
  27. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding.egg-info/requires.txt +0 -0
  28. {wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/wavenet_coding.egg-info/top_level.txt +0 -0
{wavenet_coding-0.2.0/wavenet_coding.egg-info → wavenet_coding-0.2.1}/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: wavenet-coding
3
- Version: 0.2.0
3
+ Version: 0.2.1
4
4
  Summary: WaveNet Based Low Rate Speech Coding (Parametric & Waveform)
5
5
  Home-page: https://github.com/wq2012/wavenet_coding
6
6
  Author: Quan Wang
@@ -41,7 +41,7 @@ Dynamic: summary
41
41
 
42
42
  **This is not an officially supported Google product.**
43
43
 
44
- > **Open-Source Paper Reproduction Notice**: This repository is a clean-room, open-source reproduction of the paper **W. Bastiaan Kleijn, Felicia S. C. Lim, Alejandro Luebs, Jan Skoglund, Florian Stimberg, Quan Wang, and Thomas C. Walters, *"WaveNet Based Low Rate Speech Coding,"* IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 676–680, 2018 ([arXiv:1712.01120](https://arxiv.org/abs/1712.01120))**, created after the original publication using open-source frameworks (`TensorFlow 2`, `NumPy`, `SciPy`, `SafeTensors`, `TFLite`, `GGUF`), open-source C reference codec libraries (`libcodec2`, `libspeex`, `libvo-amrwbenc`, `libopencore-amrwb`), ITU-T P.862.2 wideband `PESQ` and `STOI`, and public multi-speaker speech datasets (**LibriSpeech** and **CSTR VCTK**).
44
+ > **Open-Source Paper Reproduction Notice**: This repository is a clean-room, open-source reproduction of the paper **W. Bastiaan Kleijn, Felicia S. C. Lim, Alejandro Luebs, Jan Skoglund, Florian Stimberg, Quan Wang, and Thomas C. Walters, *"WaveNet Based Low Rate Speech Coding,"* IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 676–680, 2018 ([arXiv:1712.01120](https://arxiv.org/abs/1712.01120))**, created after the original publication using open-source frameworks (`TensorFlow 2`, `NumPy`, `SciPy`, `SafeTensors`, `TFLite`, `GGUF`), open-source C reference codec libraries (`libspeex`, `libvo-amrwbenc`, `libopencore-amrwb`) plus a pure-Python 50-bit/20ms `Codec 2` 2400 bps implementation, ITU-T P.862.2 wideband `PESQ` and `STOI`, and public multi-speaker speech datasets (**LibriSpeech** and **CSTR VCTK**).
45
45
 
46
46
  ---
47
47
 
@@ -52,9 +52,9 @@ Traditional speech coders fall into two families:
52
52
  2. **Parametric coders** (Sections 2.2 & 2.4): Extract compact vocal-tract, pitch, and energy parameters every 10–20 ms (e.g., **Codec 2** at **2.4 kb/s**) and synthesize a perceptually plausible waveform at the receiver.
53
53
 
54
54
  This package implements all four core components introduced in **Kleijn et al. (ICASSP 2018)**:
55
- - **Closed-Loop WaveNet Waveform Coding (`WW`, Section 2.3)**: Uses a causal dilated WaveNet (`q^(i)(x_i | x_0, ..., x_{i-1})`) inside a closed quantization loop over 256-level 8-bit μ-law samples (`n_i = Q(x_i)`), paired with a bit-exact 32-bit integer arithmetic coder to losslessly compress 16 kHz 8-bit μ-law speech from **128.00 kb/s** down to **80.25 kb/s** (`5.0156 bits/sample` active arithmetic coded bitrate on LibriSpeech `test-clean`) and **78.38 kb/s** (`4.8990 bits/sample` on VCTK), with a theoretical conditional entropy lower bound `H̄` of **77.69 kb/s** (`4.8556 bits/sample`) on LibriSpeech and **77.49 kb/s** (`4.8433 bits/sample`) on VCTK.
56
- - **Codec 2 Conditioned Parametric WaveNet Coding (`WP` / `W_wo` / `W_w`, Section 2.4)**: Encodes 16 kHz speech into a serialized **2.4 kb/s** bitstream (**50 bits per 20 ms block**: 36 bits LSP spectral envelope, 7 bits fundamental frequency F0, 5 bits energy, and 2 voicing flags; `36 + 7 + 5 + 2 = 50 bits / 20 ms`), unpacks the bitstream at the receiver with zero reference access, and synthesizes 16 kHz wideband speech with implicit 4–8 kHz bandwidth extension.
57
- - **Likelihood-Guided Mode Switching (`HybridModeSwitchingCodec`, Section 2.4)**: Dynamically switches between 2.4 kb/s parametric coding and closed-loop arithmetic waveform coding on 20 ms frames where the parametric negative log-likelihood exceeds a threshold (`50.09 kb/s` on LibriSpeech and `44.38 kb/s` on VCTK).
55
+ - **Closed-Loop WaveNet Waveform Coding (`WW`, Section 2.3)**: Uses a causal dilated WaveNet (`q^(i)(x_i | x_0, ..., x_{i-1})`) inside a closed quantization loop over 256-level 8-bit μ-law samples (`n_i = Q(x_i)`), paired with a bit-exact 32-bit integer arithmetic coder to losslessly compress 16 kHz 8-bit μ-law speech from **128.00 kb/s** down to **81.99 kb/s** (`5.1243 bits/sample` active arithmetic coded bitrate on LibriSpeech `test-clean`) and **77.17 kb/s** (`4.8234 bits/sample` on VCTK), with a theoretical conditional entropy lower bound `H̄` of **78.77 kb/s** (`4.9232 bits/sample`) on LibriSpeech and **77.81 kb/s** (`4.8631 bits/sample`) on VCTK.
56
+ - **Codec 2 Conditioned Parametric WaveNet Coding (`WP` / `W_wo` / `W_w`, Section 2.4)**: Encodes 16 kHz speech into a serialized **2.4 kb/s** bitstream (**43-bit** unquantized stream header + **50 bits per 20 ms block**: 36 bits LSP spectral envelope, 7 bits fundamental frequency F0, 5 bits energy, and 2 voicing flags; `36 + 7 + 5 + 2 = 50 bits / 20 ms`), unpacks the bitstream at the receiver with zero reference access, and synthesizes 16 kHz wideband speech with implicit 4–8 kHz bandwidth extension (`1.111%` high-band energy for `W_w` and `0.880%` for `W_wo` on LibriSpeech).
57
+ - **Likelihood-Guided Mode Switching (`HybridModeSwitchingCodec`, Section 2.4)**: Dynamically switches between 2.4 kb/s parametric coding and closed-loop arithmetic waveform coding on 20 ms frames where the teacher-forced parametric negative log-likelihood exceeds a threshold (`3.304 kb/s` on LibriSpeech and `2.842 kb/s` on VCTK).
58
58
  - **GE2E Speaker Verification & Triangle Discriminability (`build_speaker_verifier`, Section 3.4)**: Implements the 3-layer Projected-LSTM (`LSTMP`) speaker embedding network (128 LSTM units projected to 64 dimensions per layer over 40-channel frame-level log-Mel filterbank features) trained with Generalized End-to-End (`GE2E`) softmax loss to evaluate speaker identity preservation across 8-bit μ-law speech, speaker-disjoint Parametric WaveNet (`W_wo`), and speaker-overlapping Parametric WaveNet (`W_w`).
59
59
 
60
60
  <p align="center">
@@ -69,11 +69,11 @@ All five pretrained models are trained on real multi-speaker LibriSpeech + VCTK
69
69
 
70
70
  | Model Name | Hugging Face Repository | Task / Paper Section | Initial Eval Metric | Final Eval Metric | Exported Formats |
71
71
  | :--- | :--- | :--- | :---: | :---: | :--- |
72
- | **`wavenet_waveform_coder` (`WW`)** | [`wq2012/wavenet-waveform-coder`](https://huggingface.co/wq2012/wavenet-waveform-coder) | Closed-Loop Waveform Coder (Sec. 2.3 & 3.2) | 7.0186 bits/sample (4.8649 nats) | **4.0603 bits/sample** (2.8144 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
73
- | **`wavenet_parametric_2400_speaker_independent` (`W_wo`)** | [`wq2012/wavenet-parametric-2400-speaker-independent`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-independent) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Disjoint (Sec. 2.4 & 3.3) | 7.0185 bits/sample (4.8648 nats) | **4.0603 bits/sample** (2.8144 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
74
- | **`wavenet_parametric_2400_speaker_overlapping` (`W_w`)** | [`wq2012/wavenet-parametric-2400-speaker-overlapping`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-overlapping) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Overlapping (Sec. 3.3 & 3.4) | 7.0186 bits/sample (4.8649 nats) | **4.0603 bits/sample** (2.8144 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
75
- | **`wavenet_speaker_verifier_mulaw`** | [`wq2012/wavenet-speaker-verifier-mulaw`](https://huggingface.co/wq2012/wavenet-speaker-verifier-mulaw) | 3-Layer LSTMP GE2E Speaker Verifier on 8-bit μ-law Speech (Sec. 3.4) | 2.8332 GE2E loss (33.49% EER) | **0.3225 GE2E loss (5.81% EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
76
- | **`wavenet_speaker_verifier_coded`** | [`wq2012/wavenet-speaker-verifier-coded`](https://huggingface.co/wq2012/wavenet-speaker-verifier-coded) | 3-Layer LSTMP GE2E Speaker Verifier on 2.4 kb/s Coded Speech (Sec. 3.4) | 2.6741 GE2E loss (36.41% EER) | **0.6311 GE2E loss (11.59% EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
72
+ | **`wavenet_waveform_coder` (`WW`)** | [`wq2012/wavenet-waveform-coder`](https://huggingface.co/wq2012/wavenet-waveform-coder) | Closed-Loop Waveform Coder (Sec. 2.3 & 3.2) | 7.0186 bits/sample (4.8649 nats) | **4.2561 bits/sample** (2.9501 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
73
+ | **`wavenet_parametric_2400_speaker_independent` (`W_wo`)** | [`wq2012/wavenet-parametric-2400-speaker-independent`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-independent) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Disjoint (Sec. 2.4 & 3.3) | 7.0185 bits/sample (4.8648 nats) | **4.3116 bits/sample** (2.9886 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
74
+ | **`wavenet_parametric_2400_speaker_overlapping` (`W_w`)** | [`wq2012/wavenet-parametric-2400-speaker-overlapping`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-overlapping) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Overlapping (Sec. 3.3 & 3.4) | 7.0186 bits/sample (4.8650 nats) | **4.3116 bits/sample** (2.9886 nats, train: 4.1180) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
75
+ | **`wavenet_speaker_verifier_mulaw`** | [`wq2012/wavenet-speaker-verifier-mulaw`](https://huggingface.co/wq2012/wavenet-speaker-verifier-mulaw) | 3-Layer LSTMP GE2E Speaker Verifier on 8-bit μ-law Speech (Sec. 3.4) | 0.9159 GE2E loss (12.50% eval EER) | **0.1715 GE2E loss (2.68% eval EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
76
+ | **`wavenet_speaker_verifier_coded`** | [`wq2012/wavenet-speaker-verifier-coded`](https://huggingface.co/wq2012/wavenet-speaker-verifier-coded) | 3-Layer LSTMP GE2E Speaker Verifier on 2.4 kb/s Coded Speech (Sec. 3.4) | 1.2813 GE2E loss (15.62% eval EER) | **0.5116 GE2E loss (12.50% eval EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
77
77
 
78
78
  - **Interactive Demo Space**: [**`wq2012/wavenet-coding-demo`**](https://huggingface.co/spaces/wq2012/wavenet-coding-demo)
79
79
 
@@ -92,10 +92,10 @@ In Section 3.2 of the paper, the average conditional entropy `H̄` (Eq. 4, theor
92
92
  | Dataset / Split | Mean Conditional Entropy H̄ (bits/sample) | Theoretical Lower-Bound Bitrate (kb/s) | Cross-Entropy Rate R (bits/sample) | Cross-Entropy Bitrate (kb/s) | Active Arithmetic Coded Rate (bits/sample) | Active Arithmetic Bitrate (kb/s) | Arithmetic Coder Lossless Match Rate |
93
93
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
94
94
  | **Paper Reported (WSJ0, Sec. 3.2, 60-layer conditioned WW)** | 2.6113 (1.81 nats) | 41.78 | 2.6546 (1.84 nats) | 42.47 (~42 kb/s) | — | — | — |
95
- | **Reproduced — LibriSpeech (`test-clean`, 20-layer unconditioned WW)** | **4.8556** | **77.69** | **5.2570** (3.6438 nats) | **84.11** | **5.0156** | **80.25** | **100.0% (16/16)** |
96
- | **Reproduced — VCTK Corpus (20-layer unconditioned WW)** | **4.8433** | **77.49** | **5.0888** (3.5273 nats) | **81.42** | **4.8990** | **78.38** | **100.0% (16/16)** |
95
+ | **Reproduced — LibriSpeech (`test-clean`, 20-layer unconditioned WW)** | **4.9232** | **78.77** | **5.4089** (3.7491 nats) | **86.54** | **5.1243** | **81.99** | **100.0% (16/16)** |
96
+ | **Reproduced — VCTK Corpus (20-layer unconditioned WW)** | **4.8631** | **77.81** | **4.9755** (3.4488 nats) | **79.61** | **4.8234** | **77.17** | **100.0% (16/16)** |
97
97
 
98
- > **Architectural Scaling Note**: The original ICASSP 2018 paper trained a 60-layer conditioned WaveNet (`3` blocks × `10` layers + conditioning stack) on `32,580` WSJ0 utterances to reach `2.65 bits/sample` (`42 kb/s`). Our lightweight CPU-reproducible `20`-layer unconditioned closed-loop WaveNet (`2` blocks × `10` layers) reduces the 16 kHz 8-bit μ-law bitrate from **128.00 kb/s** to **78.38–80.25 kb/s** (`37.3%–38.8%` lossless compression savings) with **100% bit-exact reconstruction** across all 16 held-out utterances per corpus.
98
+ > **Architectural Scaling Note**: The original ICASSP 2018 paper trained a 60-layer conditioned WaveNet (`3` blocks × `10` layers + conditioning stack) on `32,580` WSJ0 utterances to reach `2.65 bits/sample` (`42 kb/s`). Our lightweight CPU-reproducible `20`-layer unconditioned closed-loop WaveNet (`2` blocks × `10` layers) reduces the 16 kHz 8-bit μ-law bitrate from **128.00 kb/s** to **77.17–81.99 kb/s** (`35.9%–39.7%` lossless compression savings) with **100% bit-exact reconstruction** across all 16 held-out utterances per corpus.
99
99
 
100
100
  #### Paper Figure 2 vs. Reproduced Instantaneous Entropy Profile
101
101
 
@@ -108,13 +108,13 @@ As observed in Section 3.2 of the paper, the instantaneous entropy `H_i` is lowe
108
108
 
109
109
  ### 2. Parametric 2.4 kb/s & Reference Codec Comparison (Section 3.3)
110
110
 
111
- Section 3.3 compares the 16 kHz Reference, Closed-Loop WaveNet Waveform Coder (`WW`), Likelihood-Guided Hybrid Mode-Switching (`WW + W_w`), **AMR-WB (`23.05 kb/s` via `libvo-amrwbenc` + `libopencore-amrwb`)**, **Parametric WaveNet with speaker overlap (`W_w, 2.4 kb/s`)**, **Parametric WaveNet without speaker overlap (`W_wo, 2.4 kb/s`)**, **MELP (`2.4 kb/s`)**, **Codec 2 (`2.4 kb/s` via `libcodec2`)**, and **Speex (`2.15 kb/s` via `libspeex`)**.
111
+ Section 3.3 compares the 16 kHz Reference, Closed-Loop WaveNet Waveform Coder (`WW`), Likelihood-Guided Hybrid Mode-Switching (`WW + W_w`), **AMR-WB (`23.05 kb/s` via `libvo-amrwbenc` + `libopencore-amrwb`)**, **Parametric WaveNet with speaker overlap (`W_w, 2.4 kb/s`)**, **Parametric WaveNet without speaker overlap (`W_wo, 2.4 kb/s`)**, **MELP (`2.4 kb/s`)**, **Codec 2 (`2.4 kb/s`)**, and **Speex (`2.15 kb/s` via `libspeex`)**.
112
112
 
113
113
  Key empirical observations on our strictly reference-free 50-bit/20ms (`2.4 kb/s`) implementation:
114
- 1. **Short-Time Objective Intelligibility (`STOI`) & Spectral Distortion (`MCD` / `LSD`)**: At **2.4 kb/s**, both Parametric WaveNet models (`W_w` and `W_wo`) achieve higher intelligibility (`STOI = 0.850` on LibriSpeech; `0.866 / 0.857` on VCTK) and significantly lower spectral envelope distortion (`MCD = 5.260–5.611 dB`, `LSD = 5.557–5.732 dB` on LibriSpeech; `MCD = 5.792–6.107 dB`, `LSD = 5.988–6.131 dB` on VCTK) than the narrow-band reference vocoders (`Codec 2`: `MCD = 8.450 / 9.775 dB`; `MELP`: `MCD = 8.111 / 9.908 dB`; `Speex`: `MCD = 7.944 / 9.161 dB`).
115
- 2. **Pitch Tracking (`F0 RMSE`)**: Conditioning the 16 kHz Parametric WaveNet decoder on the quantized 7-bit F0 contour yields lower voiced pitch error (`25.42 / 24.59 Hz` on LibriSpeech) than narrow-band `Codec 2` (`36.55 Hz`), `MELP` (`40.09 Hz`), and `Speex` (`31.14 Hz`).
116
- 3. **Implicit 4–8 kHz Bandwidth Extension**: Whereas 8 kHz narrow-band codecs (`MELP`, `Codec 2`, `Speex`) contain negligible energy above 4 kHz (`0.002%–0.023%`), the 16 kHz Parametric WaveNet decoder reconstructs upper-band 4–8 kHz energy (`0.706%` for `W_w` and `0.614%` for `W_wo` on LibriSpeech; `0.309%` and `0.291%` on VCTK) purely from the 50-bit/20ms bitstream.
117
- 4. **ITU-T P.862.2 Wideband PESQ (`MOS-LQO`)**: Because ITU-T P.862.2 `PESQ-WB` penalizes phase-asynchronous generative vocoders without high-band waveform locking, all 2.15–2.4 kb/s parametric vocoders score in the `1.35–1.69` `PESQ-WB` range (compared to `2.3–2.7` under perceptual time-warped POLQA in Table 1 of Kleijn et al.), while `AMR-WB (23.05 kb/s)` achieves `3.964 / 3.777` and lossless `WaveNet Waveform (WW)` achieves `4.171 / 3.772`.
114
+ 1. **Short-Time Objective Intelligibility (`STOI`) & Spectral Distortion (`MCD` / `LSD`)**: At **2.4 kb/s**, both Parametric WaveNet models (`W_w` and `W_wo`) achieve higher intelligibility (`STOI = 0.818` on LibriSpeech; `0.856` on VCTK) and significantly lower spectral envelope distortion (`MCD = 5.452–5.644 dB`, `LSD = 6.100–6.204 dB` on LibriSpeech; `MCD = 5.774–5.821 dB`, `LSD = 6.399–6.417 dB` on VCTK) than the narrow-band reference vocoders (`Codec 2`: `MCD = 8.383 / 9.994 dB`; `MELP`: `MCD = 8.091 / 9.629 dB`; `Speex`: `MCD = 8.055 / 9.004 dB`).
115
+ 2. **Pitch Tracking (`F0 RMSE`)**: Conditioning the 16 kHz Parametric WaveNet decoder on the quantized 7-bit F0 contour yields lower voiced pitch error (`6.872 / 6.344 Hz` on LibriSpeech; `17.184 / 17.126 Hz` on VCTK) than narrow-band `Codec 2` (`13.002 / 21.042 Hz`), `MELP` (`17.825 / 21.724 Hz`), and `Speex` (`40.598 / 40.055 Hz`).
116
+ 3. **Implicit 4–8 kHz Bandwidth Extension**: Whereas 8 kHz narrow-band codecs (`MELP`, `Codec 2`, `Speex`) contain negligible energy above 4 kHz (`0.007%–0.024%`), the 16 kHz Parametric WaveNet decoder reconstructs upper-band 4–8 kHz energy (`1.111%` for `W_w` and `0.880%` for `W_wo` on LibriSpeech; `0.545%` and `0.557%` on VCTK) purely from the 50-bit/20ms bitstream via the internal WaveNet wideband filterbank layers (`layers 1..3`).
117
+ 4. **ITU-T P.862.2 Wideband PESQ (`MOS-LQO`)**: Because ITU-T P.862.2 `PESQ-WB` penalizes phase-asynchronous generative vocoders without high-band waveform locking, all 2.15–2.4 kb/s parametric vocoders score in the `1.32–1.71` `PESQ-WB` range (compared to `2.3–2.7` under perceptual time-warped POLQA in Table 1 of Kleijn et al.), while `AMR-WB (23.05 kb/s)` achieves `4.006 / 3.809` and lossless `WaveNet Waveform (WW)` achieves `4.141 / 3.721`.
118
118
 
119
119
  #### Paper Table 1 (WSJ0 POLQA MOS-LQO, Sec. 3.3)
120
120
 
@@ -132,31 +132,31 @@ Key empirical observations on our strictly reference-free 50-bit/20ms (`2.4 kb/s
132
132
 
133
133
  | System / Codec | Bitrate (kb/s) | ITU-T P.862.2 PESQ-WB ↑ | STOI (0–1) ↑ | F0 RMSE (Hz) ↓ | MCD (dB) ↓ | LSD (dB) ↓ | SegSNR (dB) ↑ | High-Band (4–8 kHz) Energy (%) |
134
134
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
135
- | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 4.062% |
136
- | **8-bit μ-law (128 kb/s)** | 128.00 | **4.171** | **0.999** | 0.184 | 0.427 | 0.827 | 35.420 | 4.072% |
137
- | **WaveNet Waveform (`WW`)** | 84.11 (80.25 active) | **4.171** | **0.999** | 0.184 | 0.427 | 0.827 | 35.420 | 4.072% |
138
- | **Hybrid Mode-Switching (`WW + W_w`)** | 50.09 | 1.871 | 0.936 | 17.721 | 2.586 | 3.036 | 21.533 | 3.880% |
139
- | **AMR-WB (23.05 kb/s)** | 23.05 | **3.964** | **0.984** | 8.114 | 1.491 | 3.131 | 6.049 | 2.165% |
140
- | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.528 | **0.850** | 25.420 | **5.260** | **5.557** | -2.306 | **0.706%** |
141
- | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.513 | **0.850** | **24.588** | **5.611** | **5.732** | -2.297 | **0.614%** |
142
- | **MELP (2.4 kb/s)** | 2.40 | 1.686 | 0.778 | 40.092 | 8.111 | 7.895 | -2.822 | 0.020% |
143
- | **Codec 2 (2.4 kb/s)** | 2.40 | 1.597 | 0.847 | 36.549 | 8.450 | 8.067 | -2.320 | 0.005% |
144
- | **Speex (2.15 kb/s)** | 2.15 | 1.472 | 0.789 | 31.137 | 7.944 | 7.807 | -1.380 | 0.023% |
135
+ | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 5.378% |
136
+ | **8-bit μ-law (128 kb/s)** | 128.00 | **4.141** | **0.999** | 0.103 | 0.470 | 0.870 | 35.144 | 5.388% |
137
+ | **WaveNet Waveform (`WW`)** | 86.54 (81.99 active) | **4.141** | **0.999** | 0.103 | 0.470 | 0.870 | 35.144 | 5.388% |
138
+ | **Hybrid Mode-Switching (`WW + W_w`)** | 3.30 | 1.473 | **0.819** | 6.872 | **5.404** | **6.043** | -2.286 | **1.207%** |
139
+ | **AMR-WB (23.05 kb/s)** | 23.05 | **4.006** | **0.988** | 7.494 | 1.523 | 3.218 | 5.711 | 2.492% |
140
+ | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.470 | **0.818** | 6.872 | **5.452** | **6.100** | -2.466 | **1.111%** |
141
+ | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.484 | **0.818** | **6.344** | **5.644** | **6.204** | -2.446 | **0.880%** |
142
+ | **MELP (2.4 kb/s)** | 2.40 | 1.708 | 0.807 | 17.825 | 8.091 | 7.989 | -2.759 | 0.021% |
143
+ | **Codec 2 (2.4 kb/s)** | 2.40 | 1.548 | 0.793 | 13.002 | 8.383 | 8.212 | -2.187 | 0.018% |
144
+ | **Speex (2.15 kb/s)** | 2.15 | 1.477 | 0.788 | 40.598 | 8.055 | 7.971 | -1.242 | 0.024% |
145
145
 
146
146
  #### VCTK Corpus Evaluation (`pretrained_models/evaluation_vctk.json`)
147
147
 
148
148
  | System / Codec | Bitrate (kb/s) | ITU-T P.862.2 PESQ-WB ↑ | STOI (0–1) ↑ | F0 RMSE (Hz) ↓ | MCD (dB) ↓ | LSD (dB) ↓ | SegSNR (dB) ↑ | High-Band (4–8 kHz) Energy (%) |
149
149
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
150
- | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 1.774% |
151
- | **8-bit μ-law (128 kb/s)** | 128.00 | **3.772** | **0.999** | 0.107 | 0.472 | 0.810 | 34.976 | 1.783% |
152
- | **WaveNet Waveform (`WW`)** | 81.42 (78.38 active) | **3.772** | **0.999** | 0.107 | 0.472 | 0.810 | 34.976 | 1.783% |
153
- | **Hybrid Mode-Switching (`WW + W_w`)** | 44.38 | 1.430 | 0.938 | 13.676 | 3.444 | 3.820 | 19.642 | 1.289% |
154
- | **AMR-WB (23.05 kb/s)** | 23.05 | **3.777** | **0.990** | 5.780 | 1.742 | 3.522 | 5.589 | 1.113% |
155
- | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.345 | **0.866** | 33.577 | **5.792** | **5.988** | -2.276 | **0.309%** |
156
- | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.368 | **0.857** | 36.464 | **6.107** | **6.131** | -2.223 | **0.291%** |
157
- | **MELP (2.4 kb/s)** | 2.40 | 1.486 | 0.794 | 33.490 | 9.908 | 9.171 | -2.740 | 0.010% |
158
- | **Codec 2 (2.4 kb/s)** | 2.40 | 1.427 | 0.839 | 43.198 | 9.775 | 9.240 | -2.114 | 0.002% |
159
- | **Speex (2.15 kb/s)** | 2.15 | 1.491 | 0.820 | 31.791 | 9.161 | 8.986 | -1.183 | 0.017% |
150
+ | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 1.604% |
151
+ | **8-bit μ-law (128 kb/s)** | 128.00 | **3.721** | **0.999** | 0.121 | 0.552 | 0.860 | 34.748 | 1.613% |
152
+ | **WaveNet Waveform (`WW`)** | 79.61 (77.17 active) | **3.721** | **0.999** | 0.121 | 0.552 | 0.860 | 34.748 | 1.613% |
153
+ | **Hybrid Mode-Switching (`WW + W_w`)** | 2.84 | 1.318 | **0.856** | 17.184 | **5.762** | **6.404** | -2.189 | **0.551%** |
154
+ | **AMR-WB (23.05 kb/s)** | 23.05 | **3.809** | **0.990** | 9.317 | 1.719 | 3.435 | 5.398 | 1.002% |
155
+ | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.318 | **0.856** | 17.184 | **5.774** | **6.417** | -2.279 | **0.545%** |
156
+ | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.335 | **0.856** | **17.126** | **5.821** | **6.399** | -2.269 | **0.557%** |
157
+ | **MELP (2.4 kb/s)** | 2.40 | 1.460 | 0.799 | 21.724 | 9.629 | 8.989 | -2.779 | 0.011% |
158
+ | **Codec 2 (2.4 kb/s)** | 2.40 | 1.378 | 0.808 | 21.042 | 9.994 | 9.061 | -2.083 | 0.007% |
159
+ | **Speex (2.15 kb/s)** | 2.15 | 1.489 | 0.825 | 40.055 | 9.004 | 8.727 | -1.193 | 0.015% |
160
160
 
161
161
  <p align="center">
162
162
  <img src="resources/mushra.png" alt="Paper Figure 3: MUSHRA Perceptual Evaluation" width="44%"/>
@@ -165,14 +165,13 @@ Key empirical observations on our strictly reference-free 50-bit/20ms (`2.4 kb/s
165
165
 
166
166
  ### 3. Speaker Identification & Triangle Test (Section 3.4)
167
167
 
168
- Section 3.4 evaluates speaker identity preservation using two 3-layer Projected-LSTM (`LSTMP`) speaker verification networks trained with `GE2E` loss on 8-bit μ-law speech (`wavenet_speaker_verifier_mulaw`, reaching **5.81% EER** on the combined 16-speaker held-out validation set) and 2.4 kb/s coded speech (`wavenet_speaker_verifier_coded`, reaching **11.59% EER** on the combined 16-speaker held-out validation set), plus a discrete 3-AFC odd-one-out triangle test (`Original`, `W_w`, `W_wo`) measuring how often `W_wo` is selected as the odd speaker out relative to the `33.33%` random-chance baseline:
168
+ Section 3.4 evaluates speaker identity preservation using two 3-layer Projected-LSTM (`LSTMP`) speaker verification networks trained with `GE2E` loss on 8-bit μ-law speech (`wavenet_speaker_verifier_mulaw`, reaching **2.68% EER** on the held-out evaluation set) and 2.4 kb/s coded speech (`wavenet_speaker_verifier_coded`, reaching **12.50% EER** on the held-out evaluation set), plus a discrete 3-AFC odd-one-out triangle test (`Original`, `W_w`, `W_wo`) measuring how often `W_wo` is selected as the odd speaker out relative to the `33.33%` random-chance baseline:
169
169
 
170
170
  | Benchmark / Dataset | 8-bit μ-law EER (%) [95% CI] | Parametric WaveNet `W_wo` EER (%) (`V_coded`) [95% CI] | Parametric WaveNet `W_wo` EER (%) (`V_mulaw`) [95% CI] | Parametric WaveNet `W_w` EER (%) (`V_mulaw`) [95% CI] | Mean Cosine Similarity `cos(Orig, W_w)` vs `cos(Orig, W_wo)` | 3-AFC Triangle Odd-One-Out Rate (`W_wo` selected vs `33.33%` chance) |
171
171
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
172
172
  | **Paper Reported (WSJ0, Sec. 3.4)** | 8.4% | 15.8% | — | — | — | **41.0%** (chance: 33.3%) |
173
- | **Reproduced — Combined Held-Out Validation Set (16 spk, 40 utt)** | **5.81%** | **11.59%** | — | — | — | — |
174
- | **Reproduced — LibriSpeech (`test-clean`, 8 spk, 16 utt)** | **12.50%** [2.67%, 21.89%] | **22.32%** [8.48%, 32.61%] | **25.00%** [12.95%, 41.52%] | **25.00%** [13.39%, 38.40%] | **0.5918** vs **0.5684** (`Δ = +0.0233`, ABX `68.75%`) | **50.00%** (chance: 33.33%) |
175
- | **Reproduced — VCTK Corpus (8 spk, 16 utt)** | **12.50%** [3.57%, 37.51%] | **15.18%** [3.57%, 37.95%] | **24.55%** [11.15%, 37.95%] | **21.88%** [8.92%, 37.95%] | **0.6146** vs **0.6013** (`Δ = +0.0133`, ABX `56.25%`) | **43.75%** (chance: 33.33%) |
173
+ | **Reproduced — LibriSpeech (`test-clean`, 8 spk, 16 utt)** | **2.68%** [0.89%, 11.16%] | **12.50%** [1.77%, 21.44%] | **12.50%** [3.57%, 36.63%] | **12.50%** [3.57%, 34.38%] | **0.6787** vs **0.6469** (`Δ = +0.0318`, ABX `75.00%`) | **75.00%** (chance: 33.33%) |
174
+ | **Reproduced — VCTK Corpus (8 spk, 16 utt)** | **12.95%** [4.02%, 37.51%] | **23.66%** [11.60%, 37.95%] | **23.66%** [12.50%, 37.06%] | **23.66%** [9.82%, 37.50%] | **0.5678** vs **0.5584** (`Δ = +0.0094`, ABX `75.00%`) | **75.00%** (chance: 33.33%) |
176
175
 
177
176
  ---
178
177
 
{wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/README.md RENAMED
@@ -7,7 +7,7 @@
7
7
 
8
8
  **This is not an officially supported Google product.**
9
9
 
10
- > **Open-Source Paper Reproduction Notice**: This repository is a clean-room, open-source reproduction of the paper **W. Bastiaan Kleijn, Felicia S. C. Lim, Alejandro Luebs, Jan Skoglund, Florian Stimberg, Quan Wang, and Thomas C. Walters, *"WaveNet Based Low Rate Speech Coding,"* IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 676–680, 2018 ([arXiv:1712.01120](https://arxiv.org/abs/1712.01120))**, created after the original publication using open-source frameworks (`TensorFlow 2`, `NumPy`, `SciPy`, `SafeTensors`, `TFLite`, `GGUF`), open-source C reference codec libraries (`libcodec2`, `libspeex`, `libvo-amrwbenc`, `libopencore-amrwb`), ITU-T P.862.2 wideband `PESQ` and `STOI`, and public multi-speaker speech datasets (**LibriSpeech** and **CSTR VCTK**).
10
+ > **Open-Source Paper Reproduction Notice**: This repository is a clean-room, open-source reproduction of the paper **W. Bastiaan Kleijn, Felicia S. C. Lim, Alejandro Luebs, Jan Skoglund, Florian Stimberg, Quan Wang, and Thomas C. Walters, *"WaveNet Based Low Rate Speech Coding,"* IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 676–680, 2018 ([arXiv:1712.01120](https://arxiv.org/abs/1712.01120))**, created after the original publication using open-source frameworks (`TensorFlow 2`, `NumPy`, `SciPy`, `SafeTensors`, `TFLite`, `GGUF`), open-source C reference codec libraries (`libspeex`, `libvo-amrwbenc`, `libopencore-amrwb`) plus a pure-Python 50-bit/20ms `Codec 2` 2400 bps implementation, ITU-T P.862.2 wideband `PESQ` and `STOI`, and public multi-speaker speech datasets (**LibriSpeech** and **CSTR VCTK**).
11
11
 
12
12
  ---
13
13
 
@@ -18,9 +18,9 @@ Traditional speech coders fall into two families:
18
18
  2. **Parametric coders** (Sections 2.2 & 2.4): Extract compact vocal-tract, pitch, and energy parameters every 10–20 ms (e.g., **Codec 2** at **2.4 kb/s**) and synthesize a perceptually plausible waveform at the receiver.
19
19
 
20
20
  This package implements all four core components introduced in **Kleijn et al. (ICASSP 2018)**:
21
- - **Closed-Loop WaveNet Waveform Coding (`WW`, Section 2.3)**: Uses a causal dilated WaveNet (`q^(i)(x_i | x_0, ..., x_{i-1})`) inside a closed quantization loop over 256-level 8-bit μ-law samples (`n_i = Q(x_i)`), paired with a bit-exact 32-bit integer arithmetic coder to losslessly compress 16 kHz 8-bit μ-law speech from **128.00 kb/s** down to **80.25 kb/s** (`5.0156 bits/sample` active arithmetic coded bitrate on LibriSpeech `test-clean`) and **78.38 kb/s** (`4.8990 bits/sample` on VCTK), with a theoretical conditional entropy lower bound `H̄` of **77.69 kb/s** (`4.8556 bits/sample`) on LibriSpeech and **77.49 kb/s** (`4.8433 bits/sample`) on VCTK.
22
- - **Codec 2 Conditioned Parametric WaveNet Coding (`WP` / `W_wo` / `W_w`, Section 2.4)**: Encodes 16 kHz speech into a serialized **2.4 kb/s** bitstream (**50 bits per 20 ms block**: 36 bits LSP spectral envelope, 7 bits fundamental frequency F0, 5 bits energy, and 2 voicing flags; `36 + 7 + 5 + 2 = 50 bits / 20 ms`), unpacks the bitstream at the receiver with zero reference access, and synthesizes 16 kHz wideband speech with implicit 4–8 kHz bandwidth extension.
23
- - **Likelihood-Guided Mode Switching (`HybridModeSwitchingCodec`, Section 2.4)**: Dynamically switches between 2.4 kb/s parametric coding and closed-loop arithmetic waveform coding on 20 ms frames where the parametric negative log-likelihood exceeds a threshold (`50.09 kb/s` on LibriSpeech and `44.38 kb/s` on VCTK).
21
+ - **Closed-Loop WaveNet Waveform Coding (`WW`, Section 2.3)**: Uses a causal dilated WaveNet (`q^(i)(x_i | x_0, ..., x_{i-1})`) inside a closed quantization loop over 256-level 8-bit μ-law samples (`n_i = Q(x_i)`), paired with a bit-exact 32-bit integer arithmetic coder to losslessly compress 16 kHz 8-bit μ-law speech from **128.00 kb/s** down to **81.99 kb/s** (`5.1243 bits/sample` active arithmetic coded bitrate on LibriSpeech `test-clean`) and **77.17 kb/s** (`4.8234 bits/sample` on VCTK), with a theoretical conditional entropy lower bound `H̄` of **78.77 kb/s** (`4.9232 bits/sample`) on LibriSpeech and **77.81 kb/s** (`4.8631 bits/sample`) on VCTK.
22
+ - **Codec 2 Conditioned Parametric WaveNet Coding (`WP` / `W_wo` / `W_w`, Section 2.4)**: Encodes 16 kHz speech into a serialized **2.4 kb/s** bitstream (**43-bit** unquantized stream header + **50 bits per 20 ms block**: 36 bits LSP spectral envelope, 7 bits fundamental frequency F0, 5 bits energy, and 2 voicing flags; `36 + 7 + 5 + 2 = 50 bits / 20 ms`), unpacks the bitstream at the receiver with zero reference access, and synthesizes 16 kHz wideband speech with implicit 4–8 kHz bandwidth extension (`1.111%` high-band energy for `W_w` and `0.880%` for `W_wo` on LibriSpeech).
23
+ - **Likelihood-Guided Mode Switching (`HybridModeSwitchingCodec`, Section 2.4)**: Dynamically switches between 2.4 kb/s parametric coding and closed-loop arithmetic waveform coding on 20 ms frames where the teacher-forced parametric negative log-likelihood exceeds a threshold (`3.304 kb/s` on LibriSpeech and `2.842 kb/s` on VCTK).
24
24
  - **GE2E Speaker Verification & Triangle Discriminability (`build_speaker_verifier`, Section 3.4)**: Implements the 3-layer Projected-LSTM (`LSTMP`) speaker embedding network (128 LSTM units projected to 64 dimensions per layer over 40-channel frame-level log-Mel filterbank features) trained with Generalized End-to-End (`GE2E`) softmax loss to evaluate speaker identity preservation across 8-bit μ-law speech, speaker-disjoint Parametric WaveNet (`W_wo`), and speaker-overlapping Parametric WaveNet (`W_w`).
25
25
 
26
26
  <p align="center">
@@ -35,11 +35,11 @@ All five pretrained models are trained on real multi-speaker LibriSpeech + VCTK
35
35
 
36
36
  | Model Name | Hugging Face Repository | Task / Paper Section | Initial Eval Metric | Final Eval Metric | Exported Formats |
37
37
  | :--- | :--- | :--- | :---: | :---: | :--- |
38
- | **`wavenet_waveform_coder` (`WW`)** | [`wq2012/wavenet-waveform-coder`](https://huggingface.co/wq2012/wavenet-waveform-coder) | Closed-Loop Waveform Coder (Sec. 2.3 & 3.2) | 7.0186 bits/sample (4.8649 nats) | **4.0603 bits/sample** (2.8144 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
39
- | **`wavenet_parametric_2400_speaker_independent` (`W_wo`)** | [`wq2012/wavenet-parametric-2400-speaker-independent`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-independent) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Disjoint (Sec. 2.4 & 3.3) | 7.0185 bits/sample (4.8648 nats) | **4.0603 bits/sample** (2.8144 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
40
- | **`wavenet_parametric_2400_speaker_overlapping` (`W_w`)** | [`wq2012/wavenet-parametric-2400-speaker-overlapping`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-overlapping) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Overlapping (Sec. 3.3 & 3.4) | 7.0186 bits/sample (4.8649 nats) | **4.0603 bits/sample** (2.8144 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
41
- | **`wavenet_speaker_verifier_mulaw`** | [`wq2012/wavenet-speaker-verifier-mulaw`](https://huggingface.co/wq2012/wavenet-speaker-verifier-mulaw) | 3-Layer LSTMP GE2E Speaker Verifier on 8-bit μ-law Speech (Sec. 3.4) | 2.8332 GE2E loss (33.49% EER) | **0.3225 GE2E loss (5.81% EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
42
- | **`wavenet_speaker_verifier_coded`** | [`wq2012/wavenet-speaker-verifier-coded`](https://huggingface.co/wq2012/wavenet-speaker-verifier-coded) | 3-Layer LSTMP GE2E Speaker Verifier on 2.4 kb/s Coded Speech (Sec. 3.4) | 2.6741 GE2E loss (36.41% EER) | **0.6311 GE2E loss (11.59% EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
38
+ | **`wavenet_waveform_coder` (`WW`)** | [`wq2012/wavenet-waveform-coder`](https://huggingface.co/wq2012/wavenet-waveform-coder) | Closed-Loop Waveform Coder (Sec. 2.3 & 3.2) | 7.0186 bits/sample (4.8649 nats) | **4.2561 bits/sample** (2.9501 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
39
+ | **`wavenet_parametric_2400_speaker_independent` (`W_wo`)** | [`wq2012/wavenet-parametric-2400-speaker-independent`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-independent) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Disjoint (Sec. 2.4 & 3.3) | 7.0185 bits/sample (4.8648 nats) | **4.3116 bits/sample** (2.9886 nats) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
40
+ | **`wavenet_parametric_2400_speaker_overlapping` (`W_w`)** | [`wq2012/wavenet-parametric-2400-speaker-overlapping`](https://huggingface.co/wq2012/wavenet-parametric-2400-speaker-overlapping) | 2.4 kb/s Codec 2 Parametric WaveNet — Speaker-Overlapping (Sec. 3.3 & 3.4) | 7.0186 bits/sample (4.8650 nats) | **4.3116 bits/sample** (2.9886 nats, train: 4.1180) | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
41
+ | **`wavenet_speaker_verifier_mulaw`** | [`wq2012/wavenet-speaker-verifier-mulaw`](https://huggingface.co/wq2012/wavenet-speaker-verifier-mulaw) | 3-Layer LSTMP GE2E Speaker Verifier on 8-bit μ-law Speech (Sec. 3.4) | 0.9159 GE2E loss (12.50% eval EER) | **0.1715 GE2E loss (2.68% eval EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
42
+ | **`wavenet_speaker_verifier_coded`** | [`wq2012/wavenet-speaker-verifier-coded`](https://huggingface.co/wq2012/wavenet-speaker-verifier-coded) | 3-Layer LSTMP GE2E Speaker Verifier on 2.4 kb/s Coded Speech (Sec. 3.4) | 1.2813 GE2E loss (15.62% eval EER) | **0.5116 GE2E loss (12.50% eval EER)** | `.safetensors`, `saved_model/`, `.tflite` (FP32/INT8), `.gguf` (FP16/Q4_K_M) |
43
43
 
44
44
  - **Interactive Demo Space**: [**`wq2012/wavenet-coding-demo`**](https://huggingface.co/spaces/wq2012/wavenet-coding-demo)
45
45
 
@@ -58,10 +58,10 @@ In Section 3.2 of the paper, the average conditional entropy `H̄` (Eq. 4, theor
58
58
  | Dataset / Split | Mean Conditional Entropy H̄ (bits/sample) | Theoretical Lower-Bound Bitrate (kb/s) | Cross-Entropy Rate R (bits/sample) | Cross-Entropy Bitrate (kb/s) | Active Arithmetic Coded Rate (bits/sample) | Active Arithmetic Bitrate (kb/s) | Arithmetic Coder Lossless Match Rate |
59
59
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
60
60
  | **Paper Reported (WSJ0, Sec. 3.2, 60-layer conditioned WW)** | 2.6113 (1.81 nats) | 41.78 | 2.6546 (1.84 nats) | 42.47 (~42 kb/s) | — | — | — |
61
- | **Reproduced — LibriSpeech (`test-clean`, 20-layer unconditioned WW)** | **4.8556** | **77.69** | **5.2570** (3.6438 nats) | **84.11** | **5.0156** | **80.25** | **100.0% (16/16)** |
62
- | **Reproduced — VCTK Corpus (20-layer unconditioned WW)** | **4.8433** | **77.49** | **5.0888** (3.5273 nats) | **81.42** | **4.8990** | **78.38** | **100.0% (16/16)** |
61
+ | **Reproduced — LibriSpeech (`test-clean`, 20-layer unconditioned WW)** | **4.9232** | **78.77** | **5.4089** (3.7491 nats) | **86.54** | **5.1243** | **81.99** | **100.0% (16/16)** |
62
+ | **Reproduced — VCTK Corpus (20-layer unconditioned WW)** | **4.8631** | **77.81** | **4.9755** (3.4488 nats) | **79.61** | **4.8234** | **77.17** | **100.0% (16/16)** |
63
63
 
64
- > **Architectural Scaling Note**: The original ICASSP 2018 paper trained a 60-layer conditioned WaveNet (`3` blocks × `10` layers + conditioning stack) on `32,580` WSJ0 utterances to reach `2.65 bits/sample` (`42 kb/s`). Our lightweight CPU-reproducible `20`-layer unconditioned closed-loop WaveNet (`2` blocks × `10` layers) reduces the 16 kHz 8-bit μ-law bitrate from **128.00 kb/s** to **78.38–80.25 kb/s** (`37.3%–38.8%` lossless compression savings) with **100% bit-exact reconstruction** across all 16 held-out utterances per corpus.
64
+ > **Architectural Scaling Note**: The original ICASSP 2018 paper trained a 60-layer conditioned WaveNet (`3` blocks × `10` layers + conditioning stack) on `32,580` WSJ0 utterances to reach `2.65 bits/sample` (`42 kb/s`). Our lightweight CPU-reproducible `20`-layer unconditioned closed-loop WaveNet (`2` blocks × `10` layers) reduces the 16 kHz 8-bit μ-law bitrate from **128.00 kb/s** to **77.17–81.99 kb/s** (`35.9%–39.7%` lossless compression savings) with **100% bit-exact reconstruction** across all 16 held-out utterances per corpus.
65
65
 
66
66
  #### Paper Figure 2 vs. Reproduced Instantaneous Entropy Profile
67
67
 
@@ -74,13 +74,13 @@ As observed in Section 3.2 of the paper, the instantaneous entropy `H_i` is lowe
74
74
 
75
75
  ### 2. Parametric 2.4 kb/s & Reference Codec Comparison (Section 3.3)
76
76
 
77
- Section 3.3 compares the 16 kHz Reference, Closed-Loop WaveNet Waveform Coder (`WW`), Likelihood-Guided Hybrid Mode-Switching (`WW + W_w`), **AMR-WB (`23.05 kb/s` via `libvo-amrwbenc` + `libopencore-amrwb`)**, **Parametric WaveNet with speaker overlap (`W_w, 2.4 kb/s`)**, **Parametric WaveNet without speaker overlap (`W_wo, 2.4 kb/s`)**, **MELP (`2.4 kb/s`)**, **Codec 2 (`2.4 kb/s` via `libcodec2`)**, and **Speex (`2.15 kb/s` via `libspeex`)**.
77
+ Section 3.3 compares the 16 kHz Reference, Closed-Loop WaveNet Waveform Coder (`WW`), Likelihood-Guided Hybrid Mode-Switching (`WW + W_w`), **AMR-WB (`23.05 kb/s` via `libvo-amrwbenc` + `libopencore-amrwb`)**, **Parametric WaveNet with speaker overlap (`W_w, 2.4 kb/s`)**, **Parametric WaveNet without speaker overlap (`W_wo, 2.4 kb/s`)**, **MELP (`2.4 kb/s`)**, **Codec 2 (`2.4 kb/s`)**, and **Speex (`2.15 kb/s` via `libspeex`)**.
78
78
 
79
79
  Key empirical observations on our strictly reference-free 50-bit/20ms (`2.4 kb/s`) implementation:
80
- 1. **Short-Time Objective Intelligibility (`STOI`) & Spectral Distortion (`MCD` / `LSD`)**: At **2.4 kb/s**, both Parametric WaveNet models (`W_w` and `W_wo`) achieve higher intelligibility (`STOI = 0.850` on LibriSpeech; `0.866 / 0.857` on VCTK) and significantly lower spectral envelope distortion (`MCD = 5.260–5.611 dB`, `LSD = 5.557–5.732 dB` on LibriSpeech; `MCD = 5.792–6.107 dB`, `LSD = 5.988–6.131 dB` on VCTK) than the narrow-band reference vocoders (`Codec 2`: `MCD = 8.450 / 9.775 dB`; `MELP`: `MCD = 8.111 / 9.908 dB`; `Speex`: `MCD = 7.944 / 9.161 dB`).
81
- 2. **Pitch Tracking (`F0 RMSE`)**: Conditioning the 16 kHz Parametric WaveNet decoder on the quantized 7-bit F0 contour yields lower voiced pitch error (`25.42 / 24.59 Hz` on LibriSpeech) than narrow-band `Codec 2` (`36.55 Hz`), `MELP` (`40.09 Hz`), and `Speex` (`31.14 Hz`).
82
- 3. **Implicit 4–8 kHz Bandwidth Extension**: Whereas 8 kHz narrow-band codecs (`MELP`, `Codec 2`, `Speex`) contain negligible energy above 4 kHz (`0.002%–0.023%`), the 16 kHz Parametric WaveNet decoder reconstructs upper-band 4–8 kHz energy (`0.706%` for `W_w` and `0.614%` for `W_wo` on LibriSpeech; `0.309%` and `0.291%` on VCTK) purely from the 50-bit/20ms bitstream.
83
- 4. **ITU-T P.862.2 Wideband PESQ (`MOS-LQO`)**: Because ITU-T P.862.2 `PESQ-WB` penalizes phase-asynchronous generative vocoders without high-band waveform locking, all 2.15–2.4 kb/s parametric vocoders score in the `1.35–1.69` `PESQ-WB` range (compared to `2.3–2.7` under perceptual time-warped POLQA in Table 1 of Kleijn et al.), while `AMR-WB (23.05 kb/s)` achieves `3.964 / 3.777` and lossless `WaveNet Waveform (WW)` achieves `4.171 / 3.772`.
80
+ 1. **Short-Time Objective Intelligibility (`STOI`) & Spectral Distortion (`MCD` / `LSD`)**: At **2.4 kb/s**, both Parametric WaveNet models (`W_w` and `W_wo`) achieve higher intelligibility (`STOI = 0.818` on LibriSpeech; `0.856` on VCTK) and significantly lower spectral envelope distortion (`MCD = 5.452–5.644 dB`, `LSD = 6.100–6.204 dB` on LibriSpeech; `MCD = 5.774–5.821 dB`, `LSD = 6.399–6.417 dB` on VCTK) than the narrow-band reference vocoders (`Codec 2`: `MCD = 8.383 / 9.994 dB`; `MELP`: `MCD = 8.091 / 9.629 dB`; `Speex`: `MCD = 8.055 / 9.004 dB`).
81
+ 2. **Pitch Tracking (`F0 RMSE`)**: Conditioning the 16 kHz Parametric WaveNet decoder on the quantized 7-bit F0 contour yields lower voiced pitch error (`6.872 / 6.344 Hz` on LibriSpeech; `17.184 / 17.126 Hz` on VCTK) than narrow-band `Codec 2` (`13.002 / 21.042 Hz`), `MELP` (`17.825 / 21.724 Hz`), and `Speex` (`40.598 / 40.055 Hz`).
82
+ 3. **Implicit 4–8 kHz Bandwidth Extension**: Whereas 8 kHz narrow-band codecs (`MELP`, `Codec 2`, `Speex`) contain negligible energy above 4 kHz (`0.007%–0.024%`), the 16 kHz Parametric WaveNet decoder reconstructs upper-band 4–8 kHz energy (`1.111%` for `W_w` and `0.880%` for `W_wo` on LibriSpeech; `0.545%` and `0.557%` on VCTK) purely from the 50-bit/20ms bitstream via the internal WaveNet wideband filterbank layers (`layers 1..3`).
83
+ 4. **ITU-T P.862.2 Wideband PESQ (`MOS-LQO`)**: Because ITU-T P.862.2 `PESQ-WB` penalizes phase-asynchronous generative vocoders without high-band waveform locking, all 2.15–2.4 kb/s parametric vocoders score in the `1.32–1.71` `PESQ-WB` range (compared to `2.3–2.7` under perceptual time-warped POLQA in Table 1 of Kleijn et al.), while `AMR-WB (23.05 kb/s)` achieves `4.006 / 3.809` and lossless `WaveNet Waveform (WW)` achieves `4.141 / 3.721`.
84
84
 
85
85
  #### Paper Table 1 (WSJ0 POLQA MOS-LQO, Sec. 3.3)
86
86
 
@@ -98,31 +98,31 @@ Key empirical observations on our strictly reference-free 50-bit/20ms (`2.4 kb/s
98
98
 
99
99
  | System / Codec | Bitrate (kb/s) | ITU-T P.862.2 PESQ-WB ↑ | STOI (0–1) ↑ | F0 RMSE (Hz) ↓ | MCD (dB) ↓ | LSD (dB) ↓ | SegSNR (dB) ↑ | High-Band (4–8 kHz) Energy (%) |
100
100
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
101
- | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 4.062% |
102
- | **8-bit μ-law (128 kb/s)** | 128.00 | **4.171** | **0.999** | 0.184 | 0.427 | 0.827 | 35.420 | 4.072% |
103
- | **WaveNet Waveform (`WW`)** | 84.11 (80.25 active) | **4.171** | **0.999** | 0.184 | 0.427 | 0.827 | 35.420 | 4.072% |
104
- | **Hybrid Mode-Switching (`WW + W_w`)** | 50.09 | 1.871 | 0.936 | 17.721 | 2.586 | 3.036 | 21.533 | 3.880% |
105
- | **AMR-WB (23.05 kb/s)** | 23.05 | **3.964** | **0.984** | 8.114 | 1.491 | 3.131 | 6.049 | 2.165% |
106
- | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.528 | **0.850** | 25.420 | **5.260** | **5.557** | -2.306 | **0.706%** |
107
- | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.513 | **0.850** | **24.588** | **5.611** | **5.732** | -2.297 | **0.614%** |
108
- | **MELP (2.4 kb/s)** | 2.40 | 1.686 | 0.778 | 40.092 | 8.111 | 7.895 | -2.822 | 0.020% |
109
- | **Codec 2 (2.4 kb/s)** | 2.40 | 1.597 | 0.847 | 36.549 | 8.450 | 8.067 | -2.320 | 0.005% |
110
- | **Speex (2.15 kb/s)** | 2.15 | 1.472 | 0.789 | 31.137 | 7.944 | 7.807 | -1.380 | 0.023% |
101
+ | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 5.378% |
102
+ | **8-bit μ-law (128 kb/s)** | 128.00 | **4.141** | **0.999** | 0.103 | 0.470 | 0.870 | 35.144 | 5.388% |
103
+ | **WaveNet Waveform (`WW`)** | 86.54 (81.99 active) | **4.141** | **0.999** | 0.103 | 0.470 | 0.870 | 35.144 | 5.388% |
104
+ | **Hybrid Mode-Switching (`WW + W_w`)** | 3.30 | 1.473 | **0.819** | 6.872 | **5.404** | **6.043** | -2.286 | **1.207%** |
105
+ | **AMR-WB (23.05 kb/s)** | 23.05 | **4.006** | **0.988** | 7.494 | 1.523 | 3.218 | 5.711 | 2.492% |
106
+ | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.470 | **0.818** | 6.872 | **5.452** | **6.100** | -2.466 | **1.111%** |
107
+ | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.484 | **0.818** | **6.344** | **5.644** | **6.204** | -2.446 | **0.880%** |
108
+ | **MELP (2.4 kb/s)** | 2.40 | 1.708 | 0.807 | 17.825 | 8.091 | 7.989 | -2.759 | 0.021% |
109
+ | **Codec 2 (2.4 kb/s)** | 2.40 | 1.548 | 0.793 | 13.002 | 8.383 | 8.212 | -2.187 | 0.018% |
110
+ | **Speex (2.15 kb/s)** | 2.15 | 1.477 | 0.788 | 40.598 | 8.055 | 7.971 | -1.242 | 0.024% |
111
111
 
112
112
  #### VCTK Corpus Evaluation (`pretrained_models/evaluation_vctk.json`)
113
113
 
114
114
  | System / Codec | Bitrate (kb/s) | ITU-T P.862.2 PESQ-WB ↑ | STOI (0–1) ↑ | F0 RMSE (Hz) ↓ | MCD (dB) ↓ | LSD (dB) ↓ | SegSNR (dB) ↑ | High-Band (4–8 kHz) Energy (%) |
115
115
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
116
- | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 1.774% |
117
- | **8-bit μ-law (128 kb/s)** | 128.00 | **3.772** | **0.999** | 0.107 | 0.472 | 0.810 | 34.976 | 1.783% |
118
- | **WaveNet Waveform (`WW`)** | 81.42 (78.38 active) | **3.772** | **0.999** | 0.107 | 0.472 | 0.810 | 34.976 | 1.783% |
119
- | **Hybrid Mode-Switching (`WW + W_w`)** | 44.38 | 1.430 | 0.938 | 13.676 | 3.444 | 3.820 | 19.642 | 1.289% |
120
- | **AMR-WB (23.05 kb/s)** | 23.05 | **3.777** | **0.990** | 5.780 | 1.742 | 3.522 | 5.589 | 1.113% |
121
- | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.345 | **0.866** | 33.577 | **5.792** | **5.988** | -2.276 | **0.309%** |
122
- | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.368 | **0.857** | 36.464 | **6.107** | **6.131** | -2.223 | **0.291%** |
123
- | **MELP (2.4 kb/s)** | 2.40 | 1.486 | 0.794 | 33.490 | 9.908 | 9.171 | -2.740 | 0.010% |
124
- | **Codec 2 (2.4 kb/s)** | 2.40 | 1.427 | 0.839 | 43.198 | 9.775 | 9.240 | -2.114 | 0.002% |
125
- | **Speex (2.15 kb/s)** | 2.15 | 1.491 | 0.820 | 31.791 | 9.161 | 8.986 | -1.183 | 0.017% |
116
+ | **Reference (16 kHz)** | 256.00 | **4.644** | **1.000** | 0.000 | 0.000 | 0.000 | 45.000 | 1.604% |
117
+ | **8-bit μ-law (128 kb/s)** | 128.00 | **3.721** | **0.999** | 0.121 | 0.552 | 0.860 | 34.748 | 1.613% |
118
+ | **WaveNet Waveform (`WW`)** | 79.61 (77.17 active) | **3.721** | **0.999** | 0.121 | 0.552 | 0.860 | 34.748 | 1.613% |
119
+ | **Hybrid Mode-Switching (`WW + W_w`)** | 2.84 | 1.318 | **0.856** | 17.184 | **5.762** | **6.404** | -2.189 | **0.551%** |
120
+ | **AMR-WB (23.05 kb/s)** | 23.05 | **3.809** | **0.990** | 9.317 | 1.719 | 3.435 | 5.398 | 1.002% |
121
+ | **WaveNet Parametric (`W_w`, 2.4 kb/s)** | 2.40 | 1.318 | **0.856** | 17.184 | **5.774** | **6.417** | -2.279 | **0.545%** |
122
+ | **WaveNet Parametric (`W_wo`, 2.4 kb/s)** | 2.40 | 1.335 | **0.856** | **17.126** | **5.821** | **6.399** | -2.269 | **0.557%** |
123
+ | **MELP (2.4 kb/s)** | 2.40 | 1.460 | 0.799 | 21.724 | 9.629 | 8.989 | -2.779 | 0.011% |
124
+ | **Codec 2 (2.4 kb/s)** | 2.40 | 1.378 | 0.808 | 21.042 | 9.994 | 9.061 | -2.083 | 0.007% |
125
+ | **Speex (2.15 kb/s)** | 2.15 | 1.489 | 0.825 | 40.055 | 9.004 | 8.727 | -1.193 | 0.015% |
126
126
 
127
127
  <p align="center">
128
128
  <img src="resources/mushra.png" alt="Paper Figure 3: MUSHRA Perceptual Evaluation" width="44%"/>
@@ -131,14 +131,13 @@ Key empirical observations on our strictly reference-free 50-bit/20ms (`2.4 kb/s
131
131
 
132
132
  ### 3. Speaker Identification & Triangle Test (Section 3.4)
133
133
 
134
- Section 3.4 evaluates speaker identity preservation using two 3-layer Projected-LSTM (`LSTMP`) speaker verification networks trained with `GE2E` loss on 8-bit μ-law speech (`wavenet_speaker_verifier_mulaw`, reaching **5.81% EER** on the combined 16-speaker held-out validation set) and 2.4 kb/s coded speech (`wavenet_speaker_verifier_coded`, reaching **11.59% EER** on the combined 16-speaker held-out validation set), plus a discrete 3-AFC odd-one-out triangle test (`Original`, `W_w`, `W_wo`) measuring how often `W_wo` is selected as the odd speaker out relative to the `33.33%` random-chance baseline:
134
+ Section 3.4 evaluates speaker identity preservation using two 3-layer Projected-LSTM (`LSTMP`) speaker verification networks trained with `GE2E` loss on 8-bit μ-law speech (`wavenet_speaker_verifier_mulaw`, reaching **2.68% EER** on the held-out evaluation set) and 2.4 kb/s coded speech (`wavenet_speaker_verifier_coded`, reaching **12.50% EER** on the held-out evaluation set), plus a discrete 3-AFC odd-one-out triangle test (`Original`, `W_w`, `W_wo`) measuring how often `W_wo` is selected as the odd speaker out relative to the `33.33%` random-chance baseline:
135
135
 
136
136
  | Benchmark / Dataset | 8-bit μ-law EER (%) [95% CI] | Parametric WaveNet `W_wo` EER (%) (`V_coded`) [95% CI] | Parametric WaveNet `W_wo` EER (%) (`V_mulaw`) [95% CI] | Parametric WaveNet `W_w` EER (%) (`V_mulaw`) [95% CI] | Mean Cosine Similarity `cos(Orig, W_w)` vs `cos(Orig, W_wo)` | 3-AFC Triangle Odd-One-Out Rate (`W_wo` selected vs `33.33%` chance) |
137
137
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
138
138
  | **Paper Reported (WSJ0, Sec. 3.4)** | 8.4% | 15.8% | — | — | — | **41.0%** (chance: 33.3%) |
139
- | **Reproduced — Combined Held-Out Validation Set (16 spk, 40 utt)** | **5.81%** | **11.59%** | — | — | — | — |
140
- | **Reproduced — LibriSpeech (`test-clean`, 8 spk, 16 utt)** | **12.50%** [2.67%, 21.89%] | **22.32%** [8.48%, 32.61%] | **25.00%** [12.95%, 41.52%] | **25.00%** [13.39%, 38.40%] | **0.5918** vs **0.5684** (`Δ = +0.0233`, ABX `68.75%`) | **50.00%** (chance: 33.33%) |
141
- | **Reproduced — VCTK Corpus (8 spk, 16 utt)** | **12.50%** [3.57%, 37.51%] | **15.18%** [3.57%, 37.95%] | **24.55%** [11.15%, 37.95%] | **21.88%** [8.92%, 37.95%] | **0.6146** vs **0.6013** (`Δ = +0.0133`, ABX `56.25%`) | **43.75%** (chance: 33.33%) |
139
+ | **Reproduced — LibriSpeech (`test-clean`, 8 spk, 16 utt)** | **2.68%** [0.89%, 11.16%] | **12.50%** [1.77%, 21.44%] | **12.50%** [3.57%, 36.63%] | **12.50%** [3.57%, 34.38%] | **0.6787** vs **0.6469** (`Δ = +0.0318`, ABX `75.00%`) | **75.00%** (chance: 33.33%) |
140
+ | **Reproduced — VCTK Corpus (8 spk, 16 utt)** | **12.95%** [4.02%, 37.51%] | **23.66%** [11.60%, 37.95%] | **23.66%** [12.50%, 37.06%] | **23.66%** [9.82%, 37.50%] | **0.5678** vs **0.5584** (`Δ = +0.0094`, ABX `75.00%`) | **75.00%** (chance: 33.33%) |
142
141
 
143
142
  ---
144
143
 
{wavenet_coding-0.2.0 → wavenet_coding-0.2.1}/scripts/evaluate.py RENAMED
@@ -213,13 +213,32 @@ def _save_codec_comparison_plot(
213
213
  plt.close(fig)
214
214
 
215
215
 
216
+ def _round_robin_eval_pairs(
217
+ pairs: List[Any],
218
+ max_count: int,
219
+ ) -> List[Any]:
220
+ """Selects up to max_count pairs balanced across speakers."""
221
+ by_spk: Dict[str, List[str]] = {}
222
+ for spk, path in pairs:
223
+ by_spk.setdefault(str(spk), []).append(path)
224
+ ordered: List[Any] = []
225
+ max_len = max((len(v) for v in by_spk.values()), default=0)
226
+ for r in range(max_len):
227
+ for spk, paths in by_spk.items():
228
+ if r < len(paths):
229
+ ordered.append((spk, paths[r]))
230
+ return ordered[:max_count]
231
+
232
+
216
233
  def main() -> None:
217
234
  """Main entry point for evaluation."""
218
235
  args = parse_args()
219
236
  with open(args.manifest, "r", encoding="utf-8") as f:
220
237
  manifest = json.load(f)
221
238
 
222
- eval_pairs = manifest["splits"]["eval_set"][: args.max_eval_utterances]
239
+ eval_pairs = _round_robin_eval_pairs(
240
+ manifest["splits"]["eval_set"], args.max_eval_utterances
241
+ )
223
242
  ww_codec = _load_or_default_waveform_codec(args.models_dir)
224
243
  w_wo_codec = _load_or_default_parametric_codec(
225
244
  args.models_dir,
@@ -267,6 +286,18 @@ def main() -> None:
267
286
  # 1. Closed-loop WaveNet Waveform Coder (WW)
268
287
  ww_out = ww_codec.encode_and_decode(wav, run_arithmetic_coder=True)
269
288
  rm = ww_out["rate_metrics"]
289
+ if first_wav is None:
290
+ first_wav = wav
291
+ first_entropy = ww_out["instantaneous_entropy_bits"]
292
+
293
+ # 2. Parametric WaveNet Coders (W_wo and W_w) + Hybrid Mode-Switching
294
+ w_wo_out = w_wo_codec.encode_and_decode(wav, seed=42 + idx)
295
+ w_w_out = w_w_codec.encode_and_decode(wav, seed=42 + idx)
296
+ hybrid_out = hybrid_codec.encode_and_decode(
297
+ wav, seed=42 + idx, ww_out=ww_out, wp_out=w_w_out
298
+ )
299
+ rm_wo = w_wo_out["conditional_rate_metrics"]
300
+ rm_w = w_w_out["conditional_rate_metrics"]
270
301
  rate_records.append(
271
302
  {
272
303
  "mean_entropy_bits_per_sample": rm["mean_entropy_bits_per_sample"],
@@ -278,6 +309,14 @@ def main() -> None:
278
309
  rm["cross_entropy_bits_per_sample"] * np.log(2.0)
279
310
  ),
280
311
  "cross_entropy_rate_kbps": rm["cross_entropy_rate_kbps"],
312
+ "w_wo_cross_entropy_bits_per_sample": rm_wo[
313
+ "cross_entropy_bits_per_sample"
314
+ ],
315
+ "w_wo_cross_entropy_rate_kbps": rm_wo["cross_entropy_rate_kbps"],
316
+ "w_w_cross_entropy_bits_per_sample": rm_w[
317
+ "cross_entropy_bits_per_sample"
318
+ ],
319
+ "w_w_cross_entropy_rate_kbps": rm_w["cross_entropy_rate_kbps"],
281
320
  "arithmetic_coded_bits_per_sample": float(
282
321
  ww_out["arithmetic_kbps_active"] / 16.0
283
322
  ),
@@ -290,16 +329,6 @@ def main() -> None:
290
329
  ),
291
330
  }
292
331
  )
293
- if first_wav is None:
294
- first_wav = wav
295
- first_entropy = ww_out["instantaneous_entropy_bits"]
296
-
297
- # 2. Parametric WaveNet Coders (W_wo and W_w) + Hybrid Mode-Switching
298
- w_wo_out = w_wo_codec.encode_and_decode(wav, seed=42 + idx)
299
- w_w_out = w_w_codec.encode_and_decode(wav, seed=42 + idx)
300
- hybrid_out = hybrid_codec.encode_and_decode(
301
- wav, seed=42 + idx, ww_out=ww_out, wp_out=w_w_out
302
- )
303
332
 
304
333
  # 3. Reference Baseline Codecs
305
334
  frames, _ = c2_coder.encode_to_conditioning(wav)
@@ -362,28 +391,28 @@ def main() -> None:
362
391
  }
363
392
  )
364
393
 
365
- # 4. Speaker verification features (Section 3.4)
394
+ # 4. Speaker verification features (Section 3.4, 120 frames)
366
395
  mulaw_feats.append(
367
- dataset.extract_log_mel_features(mulaw_wav, max_frames=80)
396
+ dataset.extract_log_mel_features(mulaw_wav, max_frames=120)
368
397
  )
369
398
  w_wo_feats.append(
370
399
  dataset.extract_log_mel_features(
371
- w_wo_out["reconstructed_waveform"], max_frames=80
400
+ w_wo_out["reconstructed_waveform"], max_frames=120
372
401
  )
373
402
  )
374
403
  w_w_feats.append(
375
404
  dataset.extract_log_mel_features(
376
- w_w_out["reconstructed_waveform"], max_frames=80
405
+ w_w_out["reconstructed_waveform"], max_frames=120
377
406
  )
378
407
  )
379
- c2_feats.append(dataset.extract_log_mel_features(c2_wav, max_frames=80))
408
+ c2_feats.append(dataset.extract_log_mel_features(c2_wav, max_frames=120))
380
409
  melp_feats.append(
381
- dataset.extract_log_mel_features(melp_wav, max_frames=80)
410
+ dataset.extract_log_mel_features(melp_wav, max_frames=120)
382
411
  )
383
412
  speex_feats.append(
384
- dataset.extract_log_mel_features(speex_wav, max_frames=80)
413
+ dataset.extract_log_mel_features(speex_wav, max_frames=120)
385
414
  )
386
- amr_feats.append(dataset.extract_log_mel_features(amr_wav, max_frames=80))
415
+ amr_feats.append(dataset.extract_log_mel_features(amr_wav, max_frames=120))
387
416
  speaker_ids.append(spk_id)
388
417
 
389
418
  # Aggregate waveform coding rates
@@ -422,6 +451,40 @@ def main() -> None:
422
451
  float(np.mean([r["cross_entropy_rate_kbps"] for r in rate_records])),
423
452
  2,
424
453
  ),
454
+ "wavenet_parametric_w_wo_cross_entropy_bits_per_sample": round(
455
+ float(
456
+ np.mean(
457
+ [
458
+ r["w_wo_cross_entropy_bits_per_sample"]
459
+ for r in rate_records
460
+ ]
461
+ )
462
+ ),
463
+ 4,
464
+ ),
465
+ "wavenet_parametric_w_wo_cross_entropy_rate_kbps": round(
466
+ float(
467
+ np.mean([r["w_wo_cross_entropy_rate_kbps"] for r in rate_records])
468
+ ),
469
+ 2,
470
+ ),
471
+ "wavenet_parametric_w_w_cross_entropy_bits_per_sample": round(
472
+ float(
473
+ np.mean(
474
+ [
475
+ r["w_w_cross_entropy_bits_per_sample"]
476
+ for r in rate_records
477
+ ]
478
+ )
479
+ ),
480
+ 4,
481
+ ),
482
+ "wavenet_parametric_w_w_cross_entropy_rate_kbps": round(
483
+ float(
484
+ np.mean([r["w_w_cross_entropy_rate_kbps"] for r in rate_records])
485
+ ),
486
+ 2,
487
+ ),
425
488
  "arithmetic_coded_bits_per_sample": round(
426
489
  float(
427
490
  np.mean(