VieNeu-TTS v3 Turbo

VieNeu-TTS-v3-Turbo is a text-to-speech model that generates 48 kHz high-fidelity Vietnamese and bilingual English–Vietnamese speech with instant voice cloning, built-in multi-speaker default voices, and experimental emotion cues.

The model is an original architecture designed and trained from scratch by Phạm Nguyễn Ngọc Bảo on approximately 10,000 hours of English–Vietnamese speech. It uses the MOSS-Audio-Tokenizer-Nano neural audio codec and the sea-g2p grapheme-to-phoneme converter.

Key capabilities

• 48 kHz output — a substantial fidelity increase over the previous 24 kHz v2.
• Built-in default voices — ten preset voices (male and female) addressed by dedicated speaker tokens; no reference clip required for these voices.
• Instant voice cloning — clones a voice from a 3–5 second reference audio clip.
• Emotion and non-verbal cues (experimental) — supports inline tags [cười] (laugh), [thở dài] (sigh), and [hắng giọng] (clear throat).
• Batched generation — synthesises multiple chunks in one pass, batch size up to 32, including multi-speaker conversation mode.
• Bilingual code-switching — seamless transitions between Vietnamese and English within a single utterance.

Default voices

For any other voice, voice cloning with a short reference clip is used.

The model is distributed under the Apache License 2.0. A recommended temperature of 0.8 is suggested for stable results; higher values add expressiveness but may reduce stability.