Qwen-Audio 3.1 is Alibaba Qwen’s September 23, 2026 audio release: five models that Alibaba pitches as one stack for “understand, generate, interact, create.” It spans speech recognition (ASR), speech synthesis (TTS) and real-time voice chat (Realtime), plus two new models — ASR-Next for richer audio understanding and TTS-Next for unified audio generation. All five are API-only on Alibaba Cloud Model Studio and Qwen Cloud.
Alongside the models, Alibaba cut prices across the board: about 70% for TTS, about 85% for Realtime and up to 95% for ASR. That, more than any single benchmark, is the story for developers building voice products on a budget.
The five models
| Model | What it does | Key facts |
|---|---|---|
Qwen-Audio-3.1-ASR (qwen-audio-3.1-asr-flash) | Streaming speech recognition with native transcript polishing | 30 languages and 16 Chinese dialects; first-token response around 160 ms; strips filler words and repetitions |
| Qwen-Audio-3.1-ASR-Next | Audio understanding beyond words | Speaker-attributed transcripts with timestamps; understands emotion, ambient and machine sounds; sound description, event localization, audio Q&A and reasoning |
Qwen-Audio-3.1-TTS (qwen-audio-3.1-tts-flash) | Multilingual and dialect speech synthesis | Cross-language timbre transfer; instruction control of emotion, rate and style; more robust cloning under noise and reverberation |
Qwen-Audio-3.1-TTS-Next (qwen-audio-3.1-tts-next) | Unified audio generation | Speech, sound effects and ambience from text, timestamps and reference audio; podcasts, multi-speaker dialogue, cinematic soundscapes |
Qwen-Audio-3.1-Realtime (qwen-audio-3.1-realtime-plus) | Full-duplex real-time voice interaction | Listens and speaks at once; interruptions; emotion and intent understanding; tool calling |
Realtime-Plus in detail
- Full duplex: the model speaks and listens simultaneously over a streaming connection, so users can interrupt as they would on a phone call.
- Tools: function calling, or web search via
enable_search— the two cannot be enabled together — and voice cloning. - Languages: German, English, Spanish, French, Indonesian, Italian, Japanese, Korean, Portuguese, Russian and Chinese, plus many Chinese regional varieties (Cantonese, Chongqing, Northeastern and more).
- Audio formats: PCM in at 16 kHz mono, PCM out at 24 kHz mono, over WebSocket. Default voice
longanqian_v3.1, with eight new system voices. - Limits: 262,144-token context, 245,760 max input, 16,384 max output; audio history of up to 50 turns or 300 seconds; 60 RPM and 100,000 TPM.
TTS-Next limits
TTS-Next generates speech, sound effects and ambient audio in a single pass, in what press coverage describes as a language-model-plus-diffusion design. Its constraints are specific: Chinese and English only, non-streaming output (WAV, MP3 or PCM), at most 3,000 input characters, at most 240 seconds for podcasts or 120 seconds otherwise per request, and up to three reference clips (30 seconds and 10 MB each, WAV, MP3 or OGG Opus). It is rate-limited to 3 requests per second in China (Beijing).
Pricing
Per 1M tokens, from Model Studio’s list prices (promotions excluded):
| Model | International (Singapore) | China (Beijing) |
|---|---|---|
| Realtime-Plus | $0.8 text in · $6.4 audio in · $6.4 text out · $24 audio out | $0.688 · $5.501 · $5.501 · $20.628 |
| TTS-Flash | $0.23 in · $1.87 out | CNY 1.5 in · CNY 12 out |
| ASR-Flash | $0.15 in · $0.47 out | $0.113 in · $0.382 out |
| TTS-Next | — | $0.848 in · $1.696 out (CNY 6 / 12) |
For perspective, TTS output at $1.87 per 1M tokens compares with $9–$18 per 1M audio tokens for Gemini 3.8 Flash TTS — though tokens are not directly comparable across vendors, so convert to cost per minute of audio on your own samples before deciding.
Benchmarks (as published)
- Realtime (technical report, arXiv 2609.25176): Audio MultiChallenge 52.21% (+5.09 points vs 3.0), τ²-Bench Audio 82.19% (+3.59), multilingual QA 88.09%. On a half-duplex adaptation of τ-Voice, overall task success rose from 78.4% to 82.0%. Response rate to background speech fell from 0.73 to 0.13 on Full-Duplex-Bench v1.5. Alibaba notes matched 3.1 full-duplex measurements are not yet reported.
- ASR: LibriSpeech word error rate of 1.19% (clean) and 2.31% (other) in the Realtime report; dialect transcription averages 4.55% CER across 11 KeSpeech and WSYue subsets, best on 6 of them. On multi-speaker “who said what when” tests, the ASR-Next and ASR-Flash variants took 5 and 3 of 8 best-metric spots.
- TTS: Seed-TTS-Eval figures of 0.84 CER (Chinese), 1.54 WER (English) and 7.00 CER (hard set). These come from the July Qwen-Audio-3.0-TTS report that the 3.1 blog reuses, and trail some rivals on the English and hard sets — so they describe the family, not a fresh 3.1 measurement.
All results above are Alibaba’s own and not independently verified.
Independent ranking: the TTS-Plus tier
The independent number comes from Artificial Analysis’s blind-vote Speech Arena, and it needs one caveat. AA lists Qwen-Audio-3.1-TTS-Plus at #3 (Elo 1298) and Qwen-Audio-3.0-TTS-Plus at #6 (1267), both at $19.30 per 1M characters. The 3.1 entry sits behind only Eleven v4 Turbo (1330) and Eleven v4 (1326) and 24 points ahead of Gemini 3.8 Flash TTS (#5, 1274) at a similar price; the 3.0 entry is within the margin of error of Gemini. On the board where every model reads the same eight cloned voices, the 3.1 Plus entry is #2 (1185), ahead of both Eleven v4 models (1161–1166).
That entry is a character-billed Plus tier, not the token-billed TTS-Flash this page prices, and TTS-Flash is not on either board, so its quality is unmeasured. We could not find an Alibaba Model Studio page for a 3.1 Plus model ID. The documented Plus model is qwen-audio-3.0-tts-plus, listed at $0.19253 per 10,000 characters in Beijing — which is AA’s $19.30 per 1M — and $0.20 internationally. Confirm the exact model ID and price in your console before building around the ranking.
How it compares
- vs Gemini 3.8 Flash TTS: Google leads on voice design, 130 languages and two-speaker scene direction; Qwen’s TTS-Plus tier ranks above it in blind votes (#3 versus #5, and #2 versus #17 for cloned voices) at a similar $19.30 versus $16.50 per 1M characters, and bundles ASR and realtime in the same family. Gemini’s introductory price doubles in 2027.
- vs MAI-Voice-2.1: Microsoft’s pair is in public preview at $15–$22 per 1M characters with one voice across 23 languages; Qwen’s list prices look far lower on paper (tokens and characters are not directly comparable) and its Chinese dialect coverage is much deeper.
- vs Eleven v4: ElevenLabs leads blind votes by about 30 Elo points and is the more polished choice in Western markets, with 90+ languages; Qwen-Audio’s TTS-Plus tier costs $19.30 versus $40–$80 per 1M characters, and targets cost-sensitive and China-facing deployments.
- vs Qwen3.8-Omni-Flash: Omni-Flash (September 18) is a general omni-modal model that takes audio and video in but outputs text, with its own separate realtime model; Qwen-Audio 3.1 is the dedicated stack for ASR, speech generation and full-duplex voice.
Limitations
- Vendor-reported results only, with uneven documentation across the five models.
- API-only — no weights and no self-hosting.
- Regional differences. TTS-Next is listed only for Beijing, and international and China prices differ.
- Language scope. TTS-Next covers Chinese and English; Realtime covers 11 languages plus Chinese varieties.
See the audio models directory for other speech and music models.
Best for
- Voice agents and customer service that need interruption handling and tool calls (Realtime)
- Meeting transcription with speaker labels, timestamps and emotion tagging (ASR-Next)
- Low-latency voice assistants and multilingual narration (TTS-Flash)
- Podcasts, film and game soundscapes, sound effects and ambience in one pass (TTS-Next)
Pros & cons
Strengths
- A full audio stack in one family — recognition, speaker-aware understanding, TTS, unified audio generation and full-duplex voice chat — on one platform
- Realtime-Plus is full-duplex (talks and listens at once, handles interruptions), supports function calling or web search, voice cloning, and a 262,144-token context, at $24 per 1M audio output tokens
- Very low prices after a reported 70%–95% cut — TTS-Flash at $1.87 per 1M output tokens and ASR-Flash at $0.47, versus $9–$18 per 1M audio output tokens for Gemini 3.8 Flash TTS (token definitions differ by vendor)
- Strong Chinese support — 16 Chinese dialects in ASR, dozens of Chinese varieties in Realtime, and cross-language timbre transfer in TTS
Limitations
- API-only — no open weights, unlike many earlier Qwen releases
- Documentation depth varies — Realtime has a full technical report (arXiv 2609.25176), TTS reuses the July 3.0 report's numbers, and ASR-Next and TTS-Next get only brief descriptions
- TTS-Next supports Chinese and English only, non-streaming output, and 2–4 minutes per request; it is listed for the China (Beijing) region only
- Realtime-Plus speaks 11 languages (plus Chinese varieties), and web search and function calling cannot be enabled together
- Artificial Analysis ranks a character-billed TTS-Plus tier rather than the token-billed TTS-Flash covered here, and we could not find Alibaba documentation for a 3.1 Plus model ID
Frequently asked questions
What is Qwen-Audio 3.1?
Qwen-Audio 3.1 is Alibaba Qwen's audio model series, released September 23, 2026 at the Apsara Conference. It has five models — ASR, ASR-Next, TTS, TTS-Next and Realtime — that Alibaba describes as a full stack covering understanding, generation, interaction and creation. All five are available through the API only.
How much does Qwen-Audio 3.1 cost?
On Alibaba Cloud Model Studio's international (Singapore) list prices per 1M tokens, Realtime-Plus is $0.8 text input, $6.4 audio input, $6.4 text output and $24 audio output; TTS-Flash is $0.23 input and $1.87 output; ASR-Flash is $0.15 input and $0.47 output. TTS-Next is listed for China (Beijing) at $0.848 input and $1.696 output. Beijing prices for the other models are lower.
How much did Alibaba cut prices?
Alibaba says TTS prices fell about 70%, Realtime about 85% and ASR by up to 95% compared with the previous generation. These are vendor claims against its own earlier list prices, so compare your own token usage rather than relying on the percentages.
What is the difference between Realtime and TTS?
Realtime is an end-to-end, full-duplex voice conversation model that listens and speaks simultaneously, accepts interruptions and can call tools; model ID qwen-audio-3.1-realtime-plus. TTS and TTS-Flash convert text to speech, and TTS-Next generates speech, sound effects and ambience together from text, timestamps and reference audio.
How does Qwen-Audio 3.1 rank against other text-to-speech models?
On the Artificial Analysis Speech Arena (blind votes, October 2026), Qwen-Audio-3.1-TTS-Plus is
Does Qwen-Audio 3.1 support voice cloning?
Yes. Realtime-Plus supports cloned voices alongside system voices, and TTS-Flash is described as more robust at cloning under noise and reverberation. TTS-Next takes up to three reference clips of up to 30 seconds each (10 MB each).
Are Qwen-Audio 3.1 models open weights?
No. All five models are API-only through Alibaba Cloud Model Studio and Qwen Cloud.
Sources
- Qwen — Qwen-Audio-3.1 announcement (via Sina News, Sep 23, 2026)
- Alibaba Cloud Model Studio — qwen-audio-3.1-realtime-plus
- Alibaba Cloud Model Studio — qwen-audio-3.1-tts-next
- Alibaba Cloud Model Studio — qwen-audio-3.1-asr-flash
- QwenCloud — Qwen-Audio-3.1-TTS-Flash
- arXiv 2609.25176 — Qwen-Audio-3.1-Realtime: Towards Reliable Agentic Voice Interaction
- The Decoder — Alibaba launches Qwen Audio 3.1 and slashes AI audio prices
- Alibaba Cloud Model Studio — qwen-audio-3.0-tts-plus
- Artificial Analysis — Text to Speech leaderboard (provider voices)
- Artificial Analysis — Text to Speech leaderboard (controlled voice)