Model card: remove the model table
Browse files
README.md
CHANGED
|
@@ -33,22 +33,7 @@ filter) makes the waveform directly. No separate neural vocoder is needed.
|
|
| 33 |
| `tiny_voca_hindi.pt` | PyTorch checkpoint (state dict + config), after 100k training steps | 38.9 MB |
|
| 34 |
| `config.json` | the model configuration | |
|
| 35 |
|
| 36 |
-
The inference code is **not included in this repository**.
|
| 37 |
-
re-implemented.
|
| 38 |
-
|
| 39 |
-
## Model
|
| 40 |
-
|
| 41 |
-
| | |
|
| 42 |
-
|---|---|
|
| 43 |
-
| Parameters | 9,711,755 |
|
| 44 |
-
| Input | phoneme ids (70 symbols, below), a nasalization flag per phoneme, a style id |
|
| 45 |
-
| Output | 24 kHz mono audio |
|
| 46 |
-
| Text encoder | 4 transformer blocks (256 channels, 4 heads, rotary attention, depthwise conv k=7, FFN 1024) |
|
| 47 |
-
| Durations | 2 conv layers → frames per phoneme (log scale) + a pause/no-pause logit for pause-capable tokens |
|
| 48 |
-
| Prosody | 2 transformer blocks over frames → log F0, voicing, energy per 10.7 ms frame |
|
| 49 |
-
| Acoustic | 6 transformer blocks → filter cepstra and gains for the harmonic and noise branches |
|
| 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 |
|
|
|
|
| 33 |
| `tiny_voca_hindi.pt` | PyTorch checkpoint (state dict + config), after 100k training steps | 38.9 MB |
|
| 34 |
| `config.json` | the model configuration | |
|
| 35 |
|
| 36 |
+
The inference code is **not included in this repository**.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
## Training
|
| 39 |
|