Five models, one stack
Alibaba’s Qwen team released Qwen-Audio-3.1 on 23 September. It is five models rather than one: upgraded versions of the existing ASR, TTS and Realtime models, plus two new ones, ASR-Next for audio understanding and TTS-Next for audio creation.
The standard ASR model handles 30 languages and 16 Chinese dialects, with a first-character response of about 160 milliseconds. Alibaba says it also polishes transcripts as it goes, dropping filler words and repetitions.
ASR-Next goes further than transcription. It identifies which speaker said what and timestamps it, and detects emotion, ambient sound and machine noise alongside the words. TTS handles multilingual synthesis with cross-language voice transfer, and takes emotion, speed and style direction as plain text. TTS-Next pairs a language model with a diffusion approach to produce voice, sound effects and background audio in a single pass. Realtime supports speaking and listening simultaneously, with instant interruption.
The prices
The price cuts are the headline, and they are steep. Alibaba says speech recognition falls by up to 95 per cent, the realtime voice model by about 85 per cent, and text-to-speech by roughly 70 per cent. The company did not publish absolute per-unit figures with the announcement.
Percentages without a baseline are a weak disclosure, and it is fair to note that. But the direction is unmistakable, and it is the same direction Alibaba took with Qwen3.8-Omni-Flash and with its image models earlier this month.
Where this sits
Voice has become the most crowded and fastest-falling corner of the model market. Google shipped Gemini 3.8 text-to-speech with more than 2,000 voices the same week. SpaceXAI cut short-phrase transcription error by two thirds and held its price. Alibaba’s answer is to cut by an order of magnitude and widen the stack at the same time.
The two new models are the part worth watching. Most speech pipelines today are assemblies: one system to transcribe, another to diarise, another to detect sentiment, another to add sound design. ASR-Next and TTS-Next each collapse several of those steps into one model call, which changes the cost structure of a voice product more than a per-token discount does.
What Alibaba has not published is independent evaluation. Every capability claim here is the company’s own, and none of the five models has a third-party benchmark position yet. The prices are verifiable the moment anyone bills against them; the quality claims will take longer.
