Model card: remove phoneme list, GGUF notes and limits
Browse files
README.md
CHANGED
|
@@ -50,13 +50,6 @@ re-implemented.
|
|
| 50 |
| Synthesizer | band-limited pulse train at the predicted pitch + white noise, each filtered per frame (complex cepstrum, STFT overlap-add) |
|
| 51 |
| Styles | neutral, conv, happy, sad, anger, fear, surprise, disgust |
|
| 52 |
|
| 53 |
-
**Phonemes** (index = position): `<pad> <bos> <eos> <sp> <unk> , . ? ! ; : " ' ( ) — … k kʰ ɡ ɡʱ ŋ tʃ tʃʰ dʒ dʒʱ ɲ ʈ ʈʰ ɖ ɖʱ ɳ
|
| 54 |
-
t̪ t̪ʰ d̪ d̪ʱ n p pʰ b bʱ m j ɾ l ʋ ʃ ʂ s ɦ ɽ ɽʱ q x ɣ z f ʒ ə aː ɪ iː ʊ uː eː ɛː oː ɔː æ ɔ`.
|
| 55 |
-
Nasalized vowels (मैं, गाँव) use the vowel's symbol with the per-phoneme nasal flag set to 1.
|
| 56 |
-
|
| 57 |
-
The GGUF files use the architecture name `tinyvoca`: hyperparameters are stored as `tinyvoca.*` keys, along with the
|
| 58 |
-
style names, the symbol list and the phone/vowel id lists. Tensor names are the PyTorch state-dict names.
|
| 59 |
-
|
| 60 |
## Training
|
| 61 |
|
| 62 |
- **Data:** [Rasa](https://huggingface.co/datasets/ai4bharat/Rasa) by AI4Bharat, Hindi, one female speaker, with
|
|
@@ -65,14 +58,6 @@ style names, the symbol list and the phone/vowel id lists. Tensor names are the
|
|
| 65 |
- Losses: multi-resolution STFT + log-mel, plus duration, pitch and energy losses.
|
| 66 |
- Waveform critics (multi-period and multi-resolution) from step 8k, at 0.7× the reconstruction gradient.
|
| 67 |
|
| 68 |
-
## Limits
|
| 69 |
-
|
| 70 |
-
- **A faint mechanical buzz.** It is clearest on earbuds, and comes from a slightly over-smoothed (blurred) spectrum,
|
| 71 |
-
the usual weakness of models trained to predict an average spectrum.
|
| 72 |
-
- One voice only (female).
|
| 73 |
-
- Reads Hindi in Devanagari script. Numbers, dates and symbols must be written out as Hindi words before phonemization.
|
| 74 |
-
- Rare words are sometimes mispronounced.
|
| 75 |
-
|
| 76 |
## License and credits
|
| 77 |
|
| 78 |
The weights are released under **CC BY 4.0**, following the training data:
|
|
|
|
| 50 |
| Synthesizer | band-limited pulse train at the predicted pitch + white noise, each filtered per frame (complex cepstrum, STFT overlap-add) |
|
| 51 |
| Styles | neutral, conv, happy, sad, anger, fear, surprise, disgust |
|
| 52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
## Training
|
| 54 |
|
| 55 |
- **Data:** [Rasa](https://huggingface.co/datasets/ai4bharat/Rasa) by AI4Bharat, Hindi, one female speaker, with
|
|
|
|
| 58 |
- Losses: multi-resolution STFT + log-mel, plus duration, pitch and energy losses.
|
| 59 |
- Waveform critics (multi-period and multi-resolution) from step 8k, at 0.7× the reconstruction gradient.
|
| 60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
## License and credits
|
| 62 |
|
| 63 |
The weights are released under **CC BY 4.0**, following the training data:
|