Roo-Voice-NeuTTS-GGUF
Single-voice identity-locked fine-tune of NeuTTS-Air (Qwen2.5-0.5B backbone) — the custom voice model behind the Roo Voice desktop app.
Current file: v2 (checkpoint-2000) — TEST release. A v3 train with expanded lexical coverage is in progress and will replace this as the shipping model.
- Reference-free: text in, one fixed voice out. Greedy decoding (temp 0).
- Phonemization: espeak-ng en-gb, phonemizer settings must match training.
- Decode speech codes with NeuCodec (or the int8 ONNX decoder for on-device use).
- Speech-token mapping when serving via llama.cpp
/completion+return_tokens: code = token_id − 151671; stop at 151670 (<|SPEECH_GENERATION_END|>).
| File | Quant | Size | Note |
|---|---|---|---|
| roo-neutts-q4_k_m.gguf | Q4_K_M | 441 MB | app default; ear-checked |
| roo-neutts-q8_0.gguf | Q8_0 | 568 MB | headroom option; ear-checked |
Q4_0 is deliberately absent — it degraded under greedy decoding in testing.