Add INT8 tiny2.1 onnx/omograph.onnx and update docs for stress=full (v0.2.0)
Browse files- NOTICE +4 -2
- README.md +7 -9
- onnx/omograph.onnx +3 -0
NOTICE
CHANGED
|
@@ -15,8 +15,10 @@ demo.
|
|
| 15 |
|
| 16 |
The code is licensed under the Apache License 2.0 (see LICENSE).
|
| 17 |
|
| 18 |
-
The stress dictionary (stress.txt.gz)
|
| 19 |
-
(
|
|
|
|
|
|
|
| 20 |
https://github.com/Den4ikAI/ruaccent
|
| 21 |
https://huggingface.co/ruaccent/accentuator
|
| 22 |
MIT License (see src/tiny_voca_ru/frontend/data/LICENSE-ruaccent).
|
|
|
|
| 15 |
|
| 16 |
The code is licensed under the Apache License 2.0 (see LICENSE).
|
| 17 |
|
| 18 |
+
The stress dictionary (stress.txt.gz), unknown-word accent model (accent.onnx),
|
| 19 |
+
homograph tokenizer vocabulary (omograph_vocab.txt.gz), and INT8-quantized
|
| 20 |
+
homograph resolver (omograph.onnx, derived from nn/nn_omograph/tiny2.1) are
|
| 21 |
+
derived from ruaccent:
|
| 22 |
https://github.com/Den4ikAI/ruaccent
|
| 23 |
https://huggingface.co/ruaccent/accentuator
|
| 24 |
MIT License (see src/tiny_voca_ru/frontend/data/LICENSE-ruaccent).
|
README.md
CHANGED
|
@@ -29,8 +29,8 @@ A Russian adaptation of the 4.5M-parameter [`tiny-voca`](https://github.com/n4ze
|
|
| 29 |
| Parameters | 4,534,411 (~18.2 MB `safetensors`, ~21.4 MB `ONNX`) |
|
| 30 |
| Voices | `F` (female), `M` (male) |
|
| 31 |
| Output | 24 kHz mono |
|
| 32 |
-
| Frontend | Pure-Python Russian normalizer, stress (`lite` / `
|
| 33 |
-
| Weights | `model.safetensors` + `config.json`, `onnx/model.onnx` + `onnx/model.json` |
|
| 34 |
| CPU Speed (4 threads) | 115× real-time (PyTorch), 81× real-time (ONNX Runtime) |
|
| 35 |
| License | Apache-2.0 (bundled `ruaccent` stress data: MIT) |
|
| 36 |
|
|
@@ -54,8 +54,6 @@ pip install "tiny-voca-ru[onnx]"
|
|
| 54 |
# PyTorch + safetensors
|
| 55 |
pip install "tiny-voca-ru[torch]"
|
| 56 |
|
| 57 |
-
# Optional: ruaccent for homograph disambiguation
|
| 58 |
-
pip install "tiny-voca-ru[ruaccent]"
|
| 59 |
```
|
| 60 |
|
| 61 |
### Python API
|
|
@@ -74,7 +72,7 @@ result = synth.save_wav(
|
|
| 74 |
speed=1.0, # 0.25 .. 4.0
|
| 75 |
pitch_shift=0.0, # log-F0 offset (0.058 ≈ +1 semitone)
|
| 76 |
pitch_range=1.0, # 0.0 .. 3.0
|
| 77 |
-
stress="lite", # "lite" | "
|
| 78 |
)
|
| 79 |
```
|
| 80 |
|
|
@@ -118,7 +116,7 @@ sf.write("out.wav", audio, 24000)
|
|
| 118 |
Follows the [`tiny-voca`](https://github.com/n4ze3m/tiny-voca) architecture and training recipe, adapted for Russian:
|
| 119 |
|
| 120 |
- **Phoneme inventory:** 61-symbol Russian IPA set (5 special, 12 punctuation, 36 consonants with palatalization pairs, 8 vowels including reduced `ɐ` and `ə`).
|
| 121 |
-
- **Frontend (`tiny_voca_ru.frontend`):** Pure-Python Russian text normalization, stress placement (`lite` via bundled `stress.txt.gz` + `accent.onnx`, or
|
| 122 |
- **Training:** Trained from scratch for 104,000 steps (4.9 h on 1× NVIDIA L40S) on 18,517 aligned clips (21.96 h) from [`egnch/tiny-voca-ru-dataset`](https://huggingface.co/datasets/egnch/tiny-voca-ru-dataset), using `facebook/wav2vec2-lv-60-espeak-cv-ft` CTC durations during data preparation.
|
| 123 |
|
| 124 |
Custom-voice fine-tuning scripts (`tiny-voca-ru-prepare`, `tiny-voca-ru-train`) are available in the [GitHub repository](https://github.com/yilbegan/tiny-voca-ru).
|
|
@@ -132,7 +130,7 @@ Evaluated at 104k steps with the default `lite` frontend against the Qwen3-TTS t
|
|
| 132 |
| CER | test (696 clips) | 0.46% (`F` 0.43%, `M` 0.49%) | 0.31% |
|
| 133 |
| WER | test (696 clips) | 2.03% | 1.54% |
|
| 134 |
| Stress accuracy (CTC top-1) | test dict words (3,905) | 98.5% | 95.6% |
|
| 135 |
-
| Homographs (`lite` / `ruaccent`) | homographs (250) | 56.0% / 91.6% | — |
|
| 136 |
| Numbers ASR exact match | test numeric templates (139) | 90.6% | 92.1% |
|
| 137 |
| Speed (PyTorch / ONNX CPU, 4 threads) | 55.03 s audio | 115× / 81× RT | — |
|
| 138 |
| Speed (ONNX Runtime Web, WASM 1 thread) | Chromium | 16.7–21.7× RT | — |
|
|
@@ -140,7 +138,7 @@ Evaluated at 104k steps with the default `lite` frontend against the Qwen3-TTS t
|
|
| 140 |
## Limitations
|
| 141 |
|
| 142 |
- Russian only; two voices (`F`, `M`) inheriting Qwen3-TTS pronunciation traits.
|
| 143 |
-
- Default `lite` stress does not disambiguate context-dependent homographs; use `stress="
|
| 144 |
- Rule-based numeral case inflection and Latin transliteration can occasionally err on complex syntax or foreign names.
|
| 145 |
|
| 146 |
## Citation
|
|
@@ -169,6 +167,6 @@ If you use this model, please also cite the original [`tiny-voca`](https://huggi
|
|
| 169 |
## License & Attribution
|
| 170 |
|
| 171 |
- **Code & weights:** Apache-2.0 ([LICENSE](LICENSE), [NOTICE](NOTICE)).
|
| 172 |
-
- **Stress dictionary
|
| 173 |
- **Dataset:** [`egnch/tiny-voca-ru-dataset`](https://huggingface.co/datasets/egnch/tiny-voca-ru-dataset) (CC BY 4.0).
|
| 174 |
- **Acknowledgements:** [`tiny-voca`](https://github.com/n4ze3m/tiny-voca) (Apache-2.0), alignment calibration inspired by [słyszę głosy](https://github.com/damaon/slyszeglosy) (Apache-2.0), Qwen3-TTS (Apache-2.0).
|
|
|
|
| 29 |
| Parameters | 4,534,411 (~18.2 MB `safetensors`, ~21.4 MB `ONNX`) |
|
| 30 |
| Voices | `F` (female), `M` (male) |
|
| 31 |
| Output | 24 kHz mono |
|
| 32 |
+
| Frontend | Pure-Python Russian normalizer, stress (`lite` / `full`), 61-symbol IPA |
|
| 33 |
+
| Weights | `model.safetensors` + `config.json`, `onnx/model.onnx` + `onnx/model.json`, `onnx/omograph.onnx` |
|
| 34 |
| CPU Speed (4 threads) | 115× real-time (PyTorch), 81× real-time (ONNX Runtime) |
|
| 35 |
| License | Apache-2.0 (bundled `ruaccent` stress data: MIT) |
|
| 36 |
|
|
|
|
| 54 |
# PyTorch + safetensors
|
| 55 |
pip install "tiny-voca-ru[torch]"
|
| 56 |
|
|
|
|
|
|
|
| 57 |
```
|
| 58 |
|
| 59 |
### Python API
|
|
|
|
| 72 |
speed=1.0, # 0.25 .. 4.0
|
| 73 |
pitch_shift=0.0, # log-F0 offset (0.058 ≈ +1 semitone)
|
| 74 |
pitch_range=1.0, # 0.0 .. 3.0
|
| 75 |
+
stress="lite", # "lite" | "full" | "none"
|
| 76 |
)
|
| 77 |
```
|
| 78 |
|
|
|
|
| 116 |
Follows the [`tiny-voca`](https://github.com/n4ze3m/tiny-voca) architecture and training recipe, adapted for Russian:
|
| 117 |
|
| 118 |
- **Phoneme inventory:** 61-symbol Russian IPA set (5 special, 12 punctuation, 36 consonants with palatalization pairs, 8 vowels including reduced `ɐ` and `ə`).
|
| 119 |
+
- **Frontend (`tiny_voca_ru.frontend`):** Pure-Python Russian text normalization, stress placement (`lite` via bundled `stress.txt.gz` + `accent.onnx`, or `full` via 10 MB INT8 `onnx/omograph.onnx`), and rule-based G2P.
|
| 120 |
- **Training:** Trained from scratch for 104,000 steps (4.9 h on 1× NVIDIA L40S) on 18,517 aligned clips (21.96 h) from [`egnch/tiny-voca-ru-dataset`](https://huggingface.co/datasets/egnch/tiny-voca-ru-dataset), using `facebook/wav2vec2-lv-60-espeak-cv-ft` CTC durations during data preparation.
|
| 121 |
|
| 122 |
Custom-voice fine-tuning scripts (`tiny-voca-ru-prepare`, `tiny-voca-ru-train`) are available in the [GitHub repository](https://github.com/yilbegan/tiny-voca-ru).
|
|
|
|
| 130 |
| CER | test (696 clips) | 0.46% (`F` 0.43%, `M` 0.49%) | 0.31% |
|
| 131 |
| WER | test (696 clips) | 2.03% | 1.54% |
|
| 132 |
| Stress accuracy (CTC top-1) | test dict words (3,905) | 98.5% | 95.6% |
|
| 133 |
+
| Homographs (`lite` / `ruaccent turbo3.1` ref) | homographs (250) | 56.0% / 91.6% | — |
|
| 134 |
| Numbers ASR exact match | test numeric templates (139) | 90.6% | 92.1% |
|
| 135 |
| Speed (PyTorch / ONNX CPU, 4 threads) | 55.03 s audio | 115× / 81× RT | — |
|
| 136 |
| Speed (ONNX Runtime Web, WASM 1 thread) | Chromium | 16.7–21.7× RT | — |
|
|
|
|
| 138 |
## Limitations
|
| 139 |
|
| 140 |
- Russian only; two voices (`F`, `M`) inheriting Qwen3-TTS pronunciation traits.
|
| 141 |
+
- Default `lite` stress does not disambiguate context-dependent homographs; use `stress="full"` (10 MB INT8 `onnx/omograph.onnx`) or explicit `+` marks.
|
| 142 |
- Rule-based numeral case inflection and Latin transliteration can occasionally err on complex syntax or foreign names.
|
| 143 |
|
| 144 |
## Citation
|
|
|
|
| 167 |
## License & Attribution
|
| 168 |
|
| 169 |
- **Code & weights:** Apache-2.0 ([LICENSE](LICENSE), [NOTICE](NOTICE)).
|
| 170 |
+
- **Stress dictionary, `accent.onnx`, and INT8 `omograph.onnx` (`tiny2.1`):** [ruaccent](https://github.com/Den4ikAI/ruaccent) (MIT).
|
| 171 |
- **Dataset:** [`egnch/tiny-voca-ru-dataset`](https://huggingface.co/datasets/egnch/tiny-voca-ru-dataset) (CC BY 4.0).
|
| 172 |
- **Acknowledgements:** [`tiny-voca`](https://github.com/n4ze3m/tiny-voca) (Apache-2.0), alignment calibration inspired by [słyszę głosy](https://github.com/damaon/slyszeglosy) (Apache-2.0), Qwen3-TTS (Apache-2.0).
|
onnx/omograph.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:420bb1c83a5c3b5de01692fc7f73be333c6ffb181c30fe32b42294ebaa205b7d
|
| 3 |
+
size 10517401
|