On September 23, 2026, Alibaba's Qwen team officially launched the Qwen-Audio-3.1 family of voice models: five new models released in one go, covering automatic speech recognition (ASR), text-to-speech (TTS), and real-time voice interaction (Realtime), plus two brand-new "Next" models — Qwen-Audio-3.1-ASR-Next and Qwen-Audio-3.1-TTS-Next. Alongside the models came a rare round of steep voice API price cuts: TTS down about 70%, Realtime down about 85%, and ASR down up to 95%. The APIs are live on the Qwen AI platform; ASR-Next will follow later.
What each of the five models does
- Qwen-Audio-3.1-ASR: the base speech recognition model, emphasizing multilingual and dialect recognition, context understanding, and native transcription polishing.
- Qwen-Audio-3.1-TTS: the base text-to-speech model, turning text into natural-sounding voice.
- Qwen-Audio-3.1-Realtime: the real-time voice interaction model, aimed at low-latency conversational scenarios like voice assistants and phone-based customer service.
- Qwen-Audio-3.1-ASR-Next: a new model on the audio understanding side, positioned as going beyond ASR — "understanding" speech, not just transcribing it.
- Qwen-Audio-3.1-TTS-Next: the one worth a closer look. Officially an "audio creation model": given a text script, it synthesizes dialogue and ambient sound in a single pass into a complete, cinematic-grade soundscape — aimed squarely at professional creative work such as audiobooks, film dubbing, podcasts, and games.
How the price cuts add up
These are not token 10%-off gestures. TTS down about 70% means the cost of producing an hour of audiobook narration now drops to less than a third; Realtime down about 85% pulls token-hungry real-time voice conversations into affordable territory; ASR down 95% nearly flattens the marginal cost of speech transcription. For meeting notes, call quality inspection, and massive audio archives — tasks that live on volume — the math suddenly works out much better.
Why this round of cuts matters
Because it was never just about voice. On September 22, OpenAI slashed API prices for GPT-6 Sol and Luna in half (OpenAI Launches GPT-6 Sol and Luna: API Prices Cut in Half for Everyday Work (/article/2042-openai-launches-gpt-6-sol-and-luna-api-prices-cut-in-half-for-everyday-work)); the same day, Anthropic's Claude Opus 5.5 arrived roughly 40% cheaper to run than its predecessor. Now Qwen is bringing the same playbook to voice, signaling that the model price war has spread from text inference to the full voice stack. Whoever brings costs down first gets first entry into voice agents, real-time translation, and AI customer service.
Two caveats
First, TTS-Next's "single-pass complete soundscape" is a change in what a creative tool is, not just a quality bump. It shortens the production chain for audiobooks, short dramas, and podcasts — work that used to require separate dialogue recording, ambient sound design, and mixing may now be generated in one shot. But it also raises the bar for prompting: you have to know how to describe sound.
Second, ASR-Next is not live yet, and the cuts apply to API calls, not open weights. Developers who want to self-host and run their own inference get no discount this time and will have to wait for later releases.
Overall, the logic of Qwen-Audio-3.1 is clear: lock down all four links of the chain — listening, speaking, real-time chat, and creation — with a complete model lineup in one move, then smash the price barrier. Voice APIs are becoming infrastructure as cheap as text APIs, and for teams building voice products, it's time to redo the cost math.
