Qwen releases 5 audio models in one go, cuts ASR prices 95% and TTS 70%
智东西 · wechat · 2026-09-23
Qwen launched the Qwen-Audio-3.1 family: five voice models upgrading ASR, TTS and realtime interaction, plus next-gen ASR-Next (audio understanding) and TTS-Next (audio creation).
- ASR: end-to-end output with speaker labels and timestamps; 30 languages and 16 Chinese dialects; 4.55% average CER on public dialect tests, beating Doubao-ASR and Tencent Hy-ASR on 6/11 subsets; 160ms streaming first-token latency.
- TTS: one voice crosses Mandarin, Cantonese, Henan dialect, English and Japanese; emotion and pace controllable via instructions.
- Realtime: full-duplex multi-party interaction with tool calling; task success on τ-Voice up from 78.4% to 82.0%, background-speech response rate down from 73.0% to 13.0%.
- Next models: TTS-Next generates voice, sound effects and ambience with fine-grained timestamp control and 48kHz output.
Prices fell sharply: ASR -95%, Realtime -85%, TTS -70%. Qwen also unveiled Agent hardware QwenNoteA2 (magnetic phone-back assistant) and desktop robot Eva, which Realtime will connect to.
Related event: Alibaba Qwen Releases Five Qwen-Audio-3.1 Models with Massive Price Cuts(6 posts)→
More from Multimodal
- Endless screensaver of self-assembling quantum cryostats, made with Gemini 4 Carbon — hapticdata · 2026-11-12
- Adobe Firefly debuts Story Builder tool for streamlined visual storytelling — LudovicCreator · 2026-11-12
- Intangible 2.0 Launches as One AI Studio Unifying Image, Video and 3D Models — umesh_ai · 2026-11-12
- From one product image to a spy-movie ad: agent-driven video creation — hey_abusiddik · 2026-11-12
- Claude writes a Bach-style organ fugue in LilyPond with a single prompt — IgorCarron · 2026-11-12
- MiniMax H3 renders 4K character portraits twice as fast as Seedream 5.5's 2K — heyronir · 2026-11-12