add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- fa
|
| 4 |
+
license: mit
|
| 5 |
+
pipeline_tag: text-to-speech
|
| 6 |
+
base_model: rhasspy/piper-voices
|
| 7 |
+
library_name: piper
|
| 8 |
+
model_type: text-to-speech
|
| 9 |
+
tags:
|
| 10 |
+
- audio
|
| 11 |
+
- text-to-speech
|
| 12 |
+
- piper
|
| 13 |
+
- vits
|
| 14 |
+
- persian
|
| 15 |
+
- farsi
|
| 16 |
+
- fonnely
|
| 17 |
+
pretty_name: Gooya RizehPizeh v1.5
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
# Gooya RizehPizeh v1.5
|
| 21 |
+
|
| 22 |
+
Persian (Farsi) text-to-speech voice **"gooya-fa"** for [Piper](https://github.com/rhasspy/piper),
|
| 23 |
+
trained with the AvaSanj/Negara improved G2P front end. Single-speaker, 22050 Hz, `phoneme_type: text`
|
| 24 |
+
(no espeak-ng required at inference time).
|
| 25 |
+
|
| 26 |
+
## Provenance
|
| 27 |
+
|
| 28 |
+
- **Original model:** [Piper](https://github.com/rhasspy/piper) VITS, warm-started from the **Mana Persian Piper**
|
| 29 |
+
checkpoint (`epoch=6012-step=4203520.ckpt`, sdp enabled). This model is therefore a **fine-tune of Piper**,
|
| 30 |
+
continuing from a mature Persian training run rather than training from scratch.
|
| 31 |
+
- **Front end:** phonemic input produced by **Negara v7.1** G2P (grapheme-to-phoneme), with phoneme ids
|
| 32 |
+
mapped through the Mana 256-symbol inventory (157 real phonemes).
|
| 33 |
+
- **Training data:** AvaSanj **clean-core v2** — 102,584 utterances whose phoneme labels were rebuilt by the
|
| 34 |
+
OOF (out-of-fold) listener policy:
|
| 35 |
+
- `oof_listener_winner` … 48,656 (OOF AvaSanj ASR margin ≥ 0.1)
|
| 36 |
+
- `stored_audio_prompt` … 42,244 (unchanged approved prompts)
|
| 37 |
+
- `three_listener_consensus` … 11,342 (unanimous multi-listener rows)
|
| 38 |
+
- `human_override` … 41
|
| 39 |
+
- `human_reviewed_v71_overlay` … 301
|
| 40 |
+
- 28,253 rows changed vs. the stored prompt (the G2P improvement delivered by this project).
|
| 41 |
+
- **Split:** 5% validation, `num_test_examples: 0`.
|
| 42 |
+
|
| 43 |
+
## Model
|
| 44 |
+
|
| 45 |
+
- Generator parameters: **23,663,792** (~23.7 M)
|
| 46 |
+
- Architecture (Piper/VITS): `hidden_channels 192`, `filter_channels 768`, `inter_channels 192`,
|
| 47 |
+
6 flow layers, 2 attention heads, `resblock 2`, upsampling rates `[8, 8, 4]` (upsample initial channel 256),
|
| 48 |
+
`mel_channels 80`, `use_sdp true`, `num_symbols 256`, `num_speakers 1`.
|
| 49 |
+
- Vocab: 157 phoneme tokens (`text` phoneme type, Mana id map with `^`/`_`/`$` control tokens).
|
| 50 |
+
|
| 51 |
+
## Inference
|
| 52 |
+
|
| 53 |
+
```bash
|
| 54 |
+
echo "salAm olAqe aziz hAlet Cetore" | \
|
| 55 |
+
piper -m gooya-fa.onnx -c gooya-fa.onnx.json -f output.wav
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
Inference-scales baked into `gooya-fa.onnx.json`: `noise_scale 0.667`, `length_scale 1.0`, `noise_w 0.8`;
|
| 59 |
+
sample rate `22050` Hz; `espeak.voice: fa`; `phoneme_type: text`.
|
| 60 |
+
|
| 61 |
+
## Files
|
| 62 |
+
|
| 63 |
+
- `gooya-fa.onnx` — ONNX model (inference runtime)
|
| 64 |
+
- `gooya-fa.onnx.json` — Piper voice/config metadata
|
| 65 |
+
- `checkpoint/epoch=*-val_mel=*.ckpt` — PyTorch training checkpoint (resumable)
|