Roo-Voice-NeuTTS
Single-voice identity-locked fine-tune of NeuTTS-Air
(Qwen2.5-0.5B backbone) — the custom voice behind the
Roo Voice desktop app. Reference-free: text in,
one fixed voice out. Deterministic (greedy, temp 0). Phonemization: espeak-ng en-gb with
phonemizer defaults (preserve_punctuation, with_stress) — MUST match at inference.
v3 — trained fresh from base on 404 clips (~61 min) including a 123-clip lexical-coverage patch pack (s-cluster words: "class", "absolutely", …) after a word-level coverage audit of v2 found 128 common words unseen. A/B against a warm-start continuation of v2 confirmed fresh training was required: the warm start preserved v2's mispronunciations.
GGUF quants for llama.cpp serving: Roo-Voice-NeuTTS-GGUF. Decode speech codes with NeuCodec (24 kHz).