Alibaba's Qwen team released the Qwen-Audio-3.1 family of voice models on September 23, five models at once covering recognition, synthesis, realtime dialogue, and audio creation. Alongside the launch, Qwen cut prices across its voice product line: speech synthesis (TTS) down about 70%, realtime voice (Realtime) down about 85%, and speech recognition (ASR) down by up to 95%.
The five models fall into two groups. Three are upgrades to existing product lines:
- Qwen-Audio-3.1-ASR: a single model that supports 30 languages and 16 Chinese dialects. Qwen says it posts the best results on 11 dialect test subsets, with a notable improvement on harder dialects such as Wenzhounese, and a first-byte response time of about 160 milliseconds;
- Qwen-Audio-3.1-TTS: multilingual synthesis with cross-lingual voice transfer, letting users specify tone in natural language — for example, asking it to "read this in a firm, uncompromising tone";
- Qwen-Audio-3.1-Realtime: full-duplex dialogue that listens while it speaks and can be interrupted at any time, with tool-calling support; Qwen says it slows its speech down when it detects the user sounds low.
The other two are new:
- TTS-Next: a unified framework combining a language model with a diffusion model to generate voice, sound effects, and background audio in one pass, aimed at audiobooks, podcasts, games, and ads;
- ASR-Next: distinguishes speakers in multi-person audio, provides timestamps and aligned text, and can also identify emotion, ambient sound, and machine noise for audio captioning, event localization, and audio Q&A.
Dialect scorecard
Qwen's launch materials include a few dialect figures: an average character error rate of 10.38% for Chinese dialect speech recognition, and an average semantic sentence accuracy of 82.10% for dialect translation. These are Qwen's own benchmark numbers; the test set composition hasn't been fully disclosed, and there's no independent replication yet.
For developers in China, dialect recognition is a real, practical need. In customer service, government hotlines, and county-level livestream e-commerce, Mandarin recognition tends to be accurate while dialects trip systems up — one of the more common points where voice products stall in deployment. Packing 16 dialects into a single model at least removes the hassle of switching models by region.
Measured against ElevenLabs
For speech synthesis, the usual overseas benchmark is ElevenLabs. Third-party routing service OrcaRouter has compiled pricing showing the previous-generation Qwen-Audio-3.0's TTS Plus tier at roughly $27.60 per million characters and its Flash tier at about $15; ElevenLabs' Eleven v3 runs about $100 per million characters at standard pricing and about $50 on its Flash tier.
Rough math on a "roughly 70% cut" puts the 3.1 Plus tier just over $8 per million characters. That's only an estimate — the announcement gives cut percentages, not final prices, so the actual new rates will follow Alibaba Cloud's Bailian price list.
On quality, Artificial Analysis's voice arena data from August has Qwen-Audio-3.0-TTS-Plus at an Elo of 1234, against 1168 for Eleven v3. The 3.1 TTS model hasn't entered any public arena yet, so for now whether the new version is better or worse on audio quality rests only on the vendor's own claims.
Where it fits
Qwen is pitching these voice models alongside a lineup of its own products this time: the coding agent Qoder, Qwen Office, and hardware including QwenNote, A2, Eva, Qwen AI glasses, and Leqi AI glasses. Devices like glasses and voice recorders bill by the minute or per call, so an 85% to 95% cut on recognition and realtime dialogue hits hardware makers' per-device service costs most directly.
Realtime dialogue got the second-largest cut. A live voice call typically runs recognition, inference, and synthesis all at once, making it one of the pricier categories in voice products by unit cost. Whether this price cut lets voice assistants stay open to free users long-term will depend on how each integrator sets pricing next quarter.
Sources: Qwen's official announcement, IT Home, The Decoder, CocoLoop, and OrcaRouter's price roundup. Price-cut percentages and dialect test data are Qwen's own figures; legacy pricing and ElevenLabs pricing were compiled by third parties; arena scores follow Artificial Analysis's public leaderboard.