model.md

# Gooya 0.1 — Persian TTS

**Gooya 0.1** is a Persian text-to-speech model: a LoRA fine-tune of
MOSS-TTS-Nano (100M) trained on 48k Persian clips. Pronunciation comes from the
Gooya clean G2P (`Reza2kn/negara-g2p-clean`), so the model reads phonemes —
including the greeting-ezafe rule (سلام من → `salAm man`) — not raw graphemes.

- **Architecture:** global-local transformer (GPT-2 768 hidden, 12 layers + 1 local transformer) + 16-codebook audio tokenizer
- **Base:** `OpenMOSS-Team/MOSS-TTS-Nano-100M`
- **Fine-tune:** LoRA r16/α32 on 48 `transformer.h.*` linears (1.97% of params), bf16, 3 epochs
- **Data:** `Reza2kn/gooya-v7-chizzled-multispeaker` (raw Persian + audio), same-speaker ref pairs
- **Output:** 48 kHz mono WAV (via MOSS-Audio-Tokenizer-Nano)

## Usage

```bash
# 1) phonemize Persian -> char-spaced phonemes
python scripts/negara_tts_phonemize.py "سلام من رضا هستم"
#    سلام من رضا هستم	salAm man rezA hastam

# 2) synthesize (voice-clone mode)
python scripts/demo_gooya_01.py "سلام من رضا هستم" \
    --ref /path/to/reference.wav --out out.wav
```

Everything loads from Hugging Face by repo id — no local model files needed.

## Gate C (held-out, 200 rows)

6 held-out rows synthesized in two timbres:

| clip | judges |
|---|---|
| `gate_c/NN_manaref.wav` | pronunciation in base voice |
| `gate_c/NN_demoref.wav` | pronunciation + cross-voice clone |

Current verdict: pronunciation is good; manaref clips sound clearly better than
demoref clones. Voice cloning works but is imperfect.

## Known limitations

- Phoneme input only — feed raw Persian through the G2P first.
- Cloned timbre drifts on longer/lower-quality refs.
- Trained on 48k rows; pronunciation edge cases (rare homographs, foreign names)
  still benefit from the G2P rule set.

## Cite / license

Experimental research artifact. No warranty. Training recipe in
`finetune_config.json`; full pipeline in the gooya-1 repo
(`scripts/box_v7_negara.sh`).