The headline number is the price cut
Alibaba's Qwen team released a new audio lineup called Qwen-Audio-3.1 on September 23, 2026, and the detail most likely to matter to developers is not any single feature but the price. Automatic speech recognition (ASR, the technology that turns spoken audio into text) now costs up to 95 percent less through Qwen's application programming interface (API, the interface developers use to call a service from their own code). Text-to-speech (TTS, the reverse process of turning text into spoken audio) is down roughly 70 percent, and the real-time voice model is down about 85 percent.
Those are steep cuts even by the standards of a year that has already seen repeated price wars across large language model providers. For any product that transcribes calls, generates voiceovers, or runs a voice assistant at scale, the API bill for doing so just dropped sharply if that product switches to Qwen.
Five models, two of them brand new
Qwen-Audio-3.1 is built from five models rather than one. Three are upgraded versions of existing Qwen audio tools, and two are new additions built to sit alongside them.
The existing ASR model has been improved for accuracy across more languages and dialects, and it now automatically strips filler words and repeated phrases out of a transcript rather than leaving that cleanup to the developer. The existing TTS model handles multilingual speech synthesis with the ability to carry a voice across languages, and it can be steered with plain-language instructions to change emotion, speed, or style. The existing Realtime model supports full-duplex conversation, meaning it can listen and speak at the same time and be interrupted mid-sentence the way a person would be.
The two new models extend both ends of that stack. ASR-Next is aimed at understanding audio rather than just transcribing it: it adds multi-speaker identification with timestamps, so a recording of several people talking can be split out by who said what and when, along with detection of emotion and of ambient or background sounds. TTS-Next goes the other direction, pairing a language model with a diffusion-based generation approach so it can produce a voice track along with sound effects and background audio in a single pass, rather than requiring separate generation and mixing steps.
A conversational model that reads mood
The Realtime model's most distinctive claim is emotional responsiveness during a live conversation. According to Qwen's own description of the release, when the model detects that a user sounds low in mood, it responds more slowly and with more empathy, adjusting its own delivery rather than only its words.
That is a harder claim to verify from the outside than a pricing number, and Qwen has not published independent benchmark data alongside it. It is worth treating as a vendor-described feature for now rather than a proven capability, though it signals where Qwen expects real-time voice assistants to be judged going forward: not just on transcription accuracy, but on how natural the back-and-forth feels.
Availability
The five models are documented on Qwen's own blog and are live on the Qwen Cloud platform, where developers can call them through the same kind of API access Qwen already offers for its text models. No competing vendor's pricing was named in Qwen's own materials, and Alibaba has not published a company statement beyond the product blog post itself; coverage of the release has come primarily from AI-industry outlets summarizing that post rather than from an Alibaba press statement. That pattern, a product blog post doing the work a press release once did, has become routine for major Chinese AI labs shipping incremental model updates at a fast clip throughout 2026.
What it means in practice
For developers already building on Qwen's text models, adding voice through the same account is now a lower-cost decision than it was last week, and the multi-speaker and ambient-sound detection in ASR-Next in particular fills a gap that most transcription APIs still handle poorly. For teams currently paying for voice AI through a different provider, the size of the cuts, especially the 95 percent drop on ASR, is large enough to justify running a side-by-side accuracy and latency comparison before committing either way, since a lower price is only a good deal if the transcripts and voices it produces are actually usable for the task at hand.
