Alibaba’s Qwen team has introduced five Qwen Audio 3.1 models spanning transcription, speech generation and live voice interaction. ASR, TTS-Next and Realtime Plus have documented hosted endpoints. TTS Flash has a published model ID and price, but less complete integration documentation; ASR-Next has no model ID, region or price in the public developer documents reviewed by BIG CHANGE.
The Decoder reported Qwen’s advertised cuts of roughly 70% for text to speech, 85% for realtime audio and up to 95% for speech recognition. The billing units matter when checking those figures.
The big change
- What changed: Qwen now covers more of a voice product inside one hosted family, from transcription and generated speech to sound-scene production and live conversation. Its five components do not yet have equally complete developer access.
- Why it matters: For developers budgeting transcription or voice services, a minute of audio is no longer enough to price every endpoint. Token-based ASR bills include both the incoming audio and the resulting text; live conversation has four separately priced streams.
- What to watch: Alongside ASR-Next access, watch whether independent language and latency tests confirm useful savings. If they do, the wider question is whether voice-agent, transcription and dubbing providers pass those savings into customer prices, and whether rival speech APIs follow.
The five-model map
Announced model | Intended job | Documented access on September 23 | Price and practical limit |
|---|---|---|---|
Qwen Audio 3.1 ASR | Streaming, file and short-form transcription; optional speaker diarization on file and short-form endpoints | Hosted WebSocket or HTTP APIs in Singapore and Beijing. The international guide lists 30 languages and 10 Chinese dialect varieties. | Singapore ranges from $0.15/M audio-input tokens + $0.47/M text-output tokens for file/short-form to $0.93 + $0.70 for streaming/message. File transcription allows 12 hours/2GB; short-form allows 5 minutes/2GB. |
Qwen Audio 3.1 ASR-Next | Speaker attribution and timestamps plus recognition of emotion, ambient events and machine sounds | Announced by Qwen; no public API model ID, region, rate or limit found in the current Model Studio and QwenCloud documentation | Not publicly documented |
Qwen Audio 3.1 TTS | Multilingual speech, cross-language voice transfer and prompt control over delivery | | Singapore: $0.23/M text-input tokens + $1.87/M audio-output tokens. Current public limits were not specified in the reviewed 3.1 material. |
Qwen Audio 3.1 TTS-Next | Joint generation of speech, effects and background audio | Hosted non-streaming HTTPS API in Beijing, documented for Chinese and English | $0.848/M input tokens + $1.696/M output tokens. Up to 3,000 input characters; four minutes for podcast output, two minutes otherwise. |
Qwen Audio 3.1 Realtime Plus | Full-duplex voice conversation with interruption and tool calls | Hosted WebSocket API in Singapore and Beijing. The guide lists 11 languages. | Singapore: $0.80/M text input, $6.40/M audio input, $6.40/M text output and $24/M audio output. It retains up to 50 audio turns or 300 seconds of audio history. |
These are hosted API products. BIG CHANGE found no downloadable 3.1 model weights or model-weight licence. The MIT licence in QwenAudio’s GitHub repository covers the project website, so it should not be read as a licence for the models.
The language count also depends on the document. Qwen’s verified launch post says ASR covers 30 languages and 16 Chinese dialects; the international API guide names 30 languages and 10 dialect varieties. Developers should use the list attached to the endpoint they can actually call.
The 95% claim does not map to the public rate table
Alibaba’s price-change notice took effect on September 22. Its exact comparison covers qwen-audio-3.0-realtime-flash, rather than the new 3.1 Realtime Plus. The Singapore reductions are calculated as (old rate − new rate) / old rate:
Realtime Flash stream | Before (USD) | After (USD) | Reduction |
|---|---|---|---|
Text input, per million tokens | $0.45 | $0.23 | 48.89% |
Audio input, per million tokens | $4.50 | $0.93 | 79.33% |
Text output, per million tokens | $4.50 | $0.70 | 84.44% |
Output (text + audio), per million tokens | $15.00 | $1.87 | 87.53% |
The same notice moves 3.1 TTS Flash from $0.15 per 10,000 characters to separate input- and output-token rates. The current pricing page similarly lists 3.1 ASR by input and output tokens while 3.0 ASR charges per audio second with free output. A percentage across either pair depends on the text, audio duration and tokenization. The public tables reviewed here therefore do not reproduce an exact 95% ASR reduction.
Independent evidence still belongs to Qwen 3.0
No independent test of the five 3.1 models was available in the sources reviewed on launch day. Qwen’s own ASR project page reports benchmark results, but says they have not been independently rerun.
Artificial Analysis ranks Qwen Audio 3.0 TTS Plus second in its provider-voice arena (all accents and categories) at 1,259 Elo, with a 95% confidence interval of plus or minus 17 and 1,447 blind comparison samples. In a separate speech-agent arena, Qwen Audio 3.0 Realtime Plus scored 699 preference Elo with 1.54 seconds to first audio. Its participants compare hidden models in live conversations across assigned scenarios.
