86 expressive tags and sixteen languages for Alibaba's voice model

Alibaba's Qwen-Audio-3.0-TTS model integrates performance tags to dictate laughter and sighs. An advancement for dubbing in sixteen languages.

Two variants make up Qwen-Audio-3.0-TTS, the speech synthesis model from Tongyi Lab, Alibaba's research laboratory. The first, Flash, targets real-time interaction with an announced first-packet latency of around 300 ms. The second, Plus, prioritizes generation quality, natural prosody, and timbre fidelity.

Language coverage reaches sixteen languages, including Arabic, Indonesian, Portuguese, Thai, Vietnamese, Malay, and Tagalog, in addition to twenty Chinese dialectal regions. Based on its own measurements, Tongyi Lab claims the best error rates in ten of the sixteen languages and leading speaker similarity across the board. The laboratory also indicates that it holds the top spot in the text-to-speech ranking of Artificial Analysis, a third-party evaluator.

Control is achieved through two methods. A prompt written in natural language describes the expected rendering—role, emotion, style, tempo, timbre, or accent—without manual acoustic adjustment. In parallel, eighty-six tags inserted into the text act at the sentence or word level: some set a tone that runs until the next tag, such as [angry] or [whispers], while others trigger a non-verbal event at a specific point, such as a laugh, sigh, cough, or breath.

Voice cloning accepts degraded references. The training incorporated an acoustic simulation to embed denoising within the cloning pipeline, the objective being to mitigate reverberation and background noise without altering the timbre. On the architecture side, a 12.5 Hz speech tokenizer reduces the cost of autoregressive decoding. The model synthesizes up to three minutes in a single pass, comes with a library of preset voices in the sixteen languages, and is consumed via WebSocket streaming in PCM, WAV, MP3, and Opus formats. A 48 kHz output has been announced.