Qwen-Audio-3.1-ASR Technical Report

Qwen-Audio-3.1-ASR Technical Report. By Chuanmeng Bian and forty-seven others, from Alibaba Token Foundry, Alibaba Group. Published on the twentieth of October. No code repository is linked to it.

Speech recognition at its narrowest turns spoken words into text. Many uses need more than that: knowing who spoke, when, whether voices overlapped, and what else is going on in the audio besides speech. Qwen-Audio-3.1-ASR is a family of audio-language models aimed at widening that scope, building on its predecessor, Qwen-Audio-3.0-ASR.

The new family improves speech recognition and expands in two directions: speaker-diarized transcription, and general audio understanding. It keeps what the predecessor already offered. That means support for thirty languages and sixteen Chinese dialects, along with capabilities controlled by instruction: conditioning on historical context, hierarchical hotword customization, and native polishing of the transcript.

Speaker-diarized recognition produces structured output, with speaker labels, timestamps and recognized text. Speaker labels stay consistent across turns, and overlapping speech is represented. General audio understanding covers captioning of audio, speech and music, temporal grounding and sound event detection, and audio question answering with multiple-choice reasoning. Together these move the models from recognising spoken words to interpreting speakers, acoustic events and how they relate in time.

Deployment needs differ, so the family comes in variants: asynchronous file transcription, low-latency streaming, interactive voice input, and non-streaming recognition. Diarized transcription and general audio understanding are available only in certain of these variants.

Evaluations cover conventional speech recognition, multilingual and dialectal speech, speaker-diarized transcription, and general audio understanding. The report presents them as demonstrating the family's capabilities across transcription and broader audio analysis. No individual scores are stated in the summary, so how strong each capability is rests on the full report rather than on figures that can be quoted here.

The shape of this release is breadth on a shared base. One family, several deployment variants, and a set of tasks that runs from plain transcription through who said what and when, out to what the rest of the soundscape is doing. Recognition is still the core. What changes is how much of the audio the models are asked to account for.