A more context-aware transcription with Qwen-Audio 3.0

The Qwen-Audio 3.0 model uses context and keyword lists to make speech recognition more reliable. This is a game-changer for meeting minutes and reports.

The Qwen-Audio-3.0-ASR-Flash family enhances speech recognition by leveraging conversational context and specialized term lists. A developer can pass previous exchanges, add preferred words, and let the service automatically detect the language from among thirty supported ones.

Three modes cover different use cases. Qwen-Audio-3.0-ASR-Flash-Streaming processes a real-time audio stream via WebSocket, while the Flash version receives a file or a URL via HTTP. Filetrans operates asynchronously for longer recordings, with task submission followed by result retrieval once processing is complete.

In streaming mode, the service returns intermediate and then final results, accompanied by timestamps for each sentence and word. The developer can prioritize faster segmentation for voice interactions or semantic chunking better suited for meetings. The context can also be updated during transcription without interrupting the session.

The file-focused version notably adds speaker diarization, audio track selection, and the use of temporary or reusable word lists. Results are delivered as structured text with sentences, speakers, and associated timestamps.

Qwen claims to achieve a 95.36% recall rate on medical terms and 93.24% on industrial vocabulary during its internal evaluations. The company also introduces a rewriting feature designed to remove hesitations and automatically structure transcriptions.