Gooya V1 — Persian TTS
Persian text-to-speech: LoRA fine-tune of Gooya 0.1 (MOSS-TTS-Nano-100M) trained on the
102K approved corpus with v7 Negara G2P phonemes, then reinforced for long-form speech
on 2,707 clips ≥6s to remove mid-word breathing.
Training
- Stage 1 (base): Gooya 0.1 + 3-epoch LoRA on 102,084 approved clips (v7 phonemes, 30 voices)
- val_loss 4.767 → 4.708 → 4.698
- Stage 2 (long-form): continue from Stage-1, 3-epoch LoRA on 2,707 clips ≥6s
- val_loss 4.892 → 4.888 → 4.887 (long-clip eval)
- Fixes the "short-burst + mid-word breath" artifact that comes from training on short clips only
- LoRA r16/α32 on 48
transformer.h.* linears (1.97% of params), bf16
- Data:
Reza2kn/homorich-negara-gemini-tts-approved (audio + grapheme + v7 phonemes)
Inference
Phonemize Persian → char-spaced phonemes (v7 G2P) → voice-clone synthesis.
Use cool audio sampling (audio_temperature≈0.6, audio_top_p≈0.9) — the default 1.7 is too hot
and causes tremor. For long text, single-pass works up to ~12-16s after the long-form reinforcement.
Gate C / matrix
gate_c/ — held-out short clips (pre-reinforcement)matrix/ — full A/B: 2 G2P (v7/homo) × 2 TTS (gooya01/v1), v7×v1 winslong_reinforced.wav (root) — long-form demo: full 12s sentence in one pass, no mid-word breaths
