Alibaba’s Tongyi Lab Releases Qwen-Audio-3.0-TTS, a Hosted Text-to-Speech Model in Flash and Plus Tiers Across 16 Languages
Alibaba’s Tongyi Lab has launched Qwen-Audio-3.0-TTS, a production-oriented text-to-speech (TTS) system. The mannequin ships in two variants from the identical lineage. Flash targets real-time interplay. Plus targets high-quality technology. Both are delivered as hosted fashions by means of Alibaba Cloud Model Studio, not as downloadable weights.
The launch focuses on 4 issues builders hit in manufacturing: broader language protection, natural-language model management, fine-grained tag management, and robustness when the reference audio isn’t clear. Qwen-Audio-3.0-TTS-Plus additionally ranks first on the impartial Artificial Analysis Text-to-Speech leaderboard.
Two variants, one lineage
The two tiers map to totally different jobs. Flash is tuned for real-time interplay, with first-packet latency on the 300 ms degree. Plus is tuned for high-quality technology, the place naturalness and timbre constancy matter greater than pace.
The mannequin IDs are qwen-audio-3.0-tts-flash and qwen-audio-3.0-tts-plus. Both are known as over a bidirectional WebSocket streaming protocol. The API helps PCM, WAV, MP3, and Opus, with sample-rate output as much as 48 okHz. It exposes streaming enter and output, voice cloning, Voice Design, and instruction management. Alibaba gives the DashScope SDK plus uncooked WebSocket examples in Python, Java, Go, C#, PHP, and Node.js, throughout its Singapore and Beijing areas.
How the mannequin is constructed
Two design selections anchor the system.
- A 12.5 Hz low-frame-rate speech tokenizer reduces autoregressive decoding price whereas retaining content material and speaker data. A decrease body charge means fewer tokens per second of audio, which cuts inference latency.
- A five-stage progressive coaching paradigm coordinates the language mannequin (LM) and flow-matching (FM) elements. The phases are impartial LM and FM pretraining, joint coaching with high-quality information annealing, LM reinforcement studying, FM robustness coaching, and FM reinforcement studying. The analysis group reviews this pipeline improves content material consistency, prosodic naturalness, voice constancy, perceptual high quality, and robustness.
The mannequin additionally handles one-pass long-form synthesis as much as 3 minutes, arduous text-normalization circumstances, and vocoder super-resolution for 48 okHz output.
Key Takeaways
- Qwen-Audio-3.0-TTS is a hosted TTS mannequin in two tiers: Flash (~300 ms first-packet, real-time) and Plus (quality-first).
- Plus ranks #1 on the Artificial Analysis area at ~1,236 Elo, priced at ~$27.59 per 1M characters, however solely ~16 chars/sec throughput.
- Coverage spans 16 languages and 20 Chinese dialects, with greatest WER/CER in 10 of 16 languages and high speaker similarity on Plus.
- Control comes two methods: free-style natural-language directions and 86 fine-grained inline tags for non-verbal element.
- It is distinct from the open-weight Qwen3-TTS (Apache-2.0) line; the three.0 mannequin is API-only through Alibaba Cloud Model Studio.
Sources: Qwen-Audio-3.0-TTS release blog · Tongyi Lab announcement · Alibaba Cloud Model Studio real-time TTS docs · Artificial Analysis leaderboard · Tongyi Lab on X
The submit Alibaba’s Tongyi Lab Releases Qwen-Audio-3.0-TTS, a Hosted Text-to-Speech Model in Flash and Plus Tiers Across 16 Languages appeared first on MarkTechPost.
