Qwen-Audio-3.1-ASR: 30 Languages, 16 Dialects, and Diarization in a Single Model

Qwen has published a technical report for Qwen-Audio-3.1-ASR, a family of audio-language models that builds on the Qwen-Audio-3.0-ASR line. The basic parameters are unchanged: 30 languages and 16 Chinese dialects. The range of tasks, however, has grown noticeably. The model now labels speech by speaker with timestamps and answers questions about audio content, including music and background sounds. In effect, an ASR model is turning into a general-purpose audio analyst.

Diarization that keeps speakers straight

The most practical update is transcription with speaker separation. The output is not a block of plain text but a structure: speaker label, time interval, recognized phrase. The authors single out two problems on which such systems usually break down. The first is label consistency across turns: Speaker 1 at the start of a recording must still be Speaker 1 forty minutes later. The second is overlapping speech, when two people talk at once. For calls, interviews, and call centers, this is a core pain point: if the model confuses people or merges interrupting turns into one, the transcript loses its value.

The report evaluates this scenario too. Its benchmarks are split into classic ASR, multilingual and dialect speech, diarized transcription, and general audio understanding.

Control through instructions

The second line of development is that the model's behavior can be tuned with text prompts. The family retains three mechanisms:

• Conditioning on historical context. The model takes earlier parts of the conversation into account, which helps with homonyms and proper names.
• Hierarchical hotwords. Glossaries of terms, names, and titles can be set with priorities rather than as a flat list.
• Native transcript polishing. The model itself removes disfluencies, adds punctuation, and turns the text into a readable form without separate post-processing.

All of this is controlled through instructions, not fine-tuning. For companies with specialized vocabulary, such as medical, legal, or financial, this approach is considerably cheaper than training a custom model for each domain.

General audio understanding covers:

• description of audio, speech, and music (captioning);
• temporal grounding and detection of sound events;
• question answering about audio;
• multiple-choice questions that require reasoning.

This shifts the range of applications from "what was said" to "who said it, what was sounding around them, and at what moment."

Four modes for different scenarios

The family is split into variants for different deployment conditions: asynchronous file transcription, low-latency streaming, interactive voice input, and non-streaming recognition. Diarization and general audio understanding are available only in some variants, not all of them. That makes sense: minimal-latency streaming and deep analysis of a long recording call for different trade-offs between speed and quality.

Who will benefit first

The biggest winners are those who work with long, multilingual recordings: meeting transcription services, media outlets with interview archives, contact centers, and video content platforms. A single API in place of a stack of a recognizer, a diarizer, and a separate model for sound analysis simplifies the architecture and reduces the number of failure points.

Support for 16 Chinese dialects makes the family especially interesting for the Chinese market and diaspora communities, where standard ASR systems handle regional speech noticeably worse. For developers of voice interfaces, the low-latency variant with interactive input covers the assistant use case.

Before adopting it, check the specific variants: not all of them offer diarization. It is best to test on your own recordings with interruptions and noise, which is where such models usually fall short of their claimed quality.