Qwen-Audio-3.1-ASR Adds Speaker-Aware Transcription

Alibaba's Qwen team has introduced Qwen-Audio-3.1-ASR, an audio-language model focused on automatic speech recognition that pairs transcription with speaker diarization across 30 languages. The release is documented in a technical report on Hugging Face and carries a permissive Apache 2.0 license.

The headline capability is combining who-spoke-when with what-was-said. Rather than producing a flat transcript, the model attributes speech to distinct speakers — a feature that is often bolted on separately in traditional ASR pipelines but is increasingly expected in a single system.

Why it matters

Speaker-diarized, multilingual ASR is the connective tissue for a lot of practical tooling, and having it in an openly licensed model lowers the barrier for teams building on top of it.

• Coverage spans 30 languages, widening its reach beyond English-centric systems.
• Built-in diarization suits meeting notes, interviews, and call transcription.
• Apache 2.0 allows commercial use and modification without heavy restrictions.

The technical report does not detail parameter counts or context limits, so some specifics remain to be filled in. Still, as part of the broader Qwen audio-language effort, the 3.1-ASR line signals continued investment in open speech models that handle real-world, multi-speaker audio rather than clean single-voice recordings.