Qwen-Audio-3.1 brings five voice models for transcription, speech and live conversation
Alibaba's Qwen team has refreshed its audio model family with Qwen-Audio-3.1: five models that transcribe speech, turn text into speech and hold live conversations. The models are available by API on Qwen Cloud.
Alibaba's Qwen team has announced Qwen-Audio-3.1, the new version of its audio model family. It has five models covering four jobs in one stack: understanding, generation, conversation and creation. The models are available by API through Qwen Cloud and Alibaba Cloud Model Studio. Qwen also announced large price cuts with the release.
Five models
- ASR (speech to text): Transcribes short speech in many languages. According to its model page, it adds punctuation, cleans up text and separates who is speaking in multi-person conversations. It also recognises regional Chinese dialects.
- ASR-Next (new): A deeper audio-understanding model. According to Qwen, it separates speakers, adds timestamps and recognises emotion and ambient sounds.
- TTS (text to speech): The model that turns text into natural speech, upgraded in this release.
- TTS-Next (new): Generates a complete audio track, not just speech. From text, timing and reference audio, it combines speech, sound effects and ambient sound in one pass. It is aimed at podcasts, multi-speaker dialogue, short video, games and film audio.
- Realtime (live conversation): A real-time voice model that can listen and speak at the same time. It handles interruptions, waits for the user to finish, and copes better with noise and multi-person conversations. It speaks several languages and can call tools to connect to company knowledge bases and business systems. Qwen suggests it for customer service and voice assistants.
How to use it
The models are called with an API key from their pages on Qwen Cloud; the live conversation model runs over WebSocket. The model pages include sample code. Usage is billed, and current prices are listed on Qwen Cloud.
Things to keep in mind
- Qwen-Audio-3.1 is offered as a cloud API. The model files cannot be downloaded or run on your own servers. Qwen3-TTS, which Qwen released earlier as an open model, is a separate series.
- According to Alibaba Cloud's docs, TTS-Next currently supports only Chinese and English.
- For a fuller picture, follow evaluations from independent researchers as well.
Sources: Qwen announcement, Qwen Cloud: Qwen-Audio-3.1-Realtime, Qwen Cloud: Qwen-Audio-3.1-ASR, Alibaba Cloud: Qwen-Audio-3.1-TTS-Next.
