01
Streaming and file transcription are separate paths
Alibaba documents qwen-audio-3.1-asr-flash-streaming for real-time WebSocket recognition and qwen-audio-3.1-asr-flash-filetrans for asynchronous recorded-file transcription. Choosing the endpoint by workload avoids mixing incompatible limits and features.
Use streaming for live captions, online meetings and voice applications. Use file transcription for completed recordings such as interviews, calls and long meetings where asynchronous processing and diarization are useful.
02
Long files, streaming and diarization
The current non-real-time guide documents file transcription up to 12 hours or 2 GB. When speaker diarization is enabled, Alibaba advises keeping audio within two hours because longer files may fail or time out.
The streaming model is documented for real-time recognition, but production clients still need reconnect, timeout and backoff handling. A model table's duration category should not be interpreted as a guarantee that any network connection can remain open indefinitely.
03
Languages, hotwords and context enhancement
Qwen-Audio-3.1 ASR supports a broad multilingual set and multiple Chinese dialects. Exact language support should be checked against the specific ASR variant instead of inherited from another Qwen audio model.
Alibaba documents hotwords for domain-specific names and context enhancement for recognition of proper nouns. These are useful in interviews, customer calls and technical meetings where generic speech models often miss product and organization names.
04
Asynchronous file workflow and output
The long-file workflow is task based: submit the file, receive a task ID, poll or wait for completion, then download the result JSON from the returned URL. High-concurrency systems should follow Alibaba's callback guidance rather than polling aggressively.
Timestamps and diarization are endpoint-specific features. If your only goal is a transcript, ASR is the direct tool; if the application needs a spoken response as part of a live interaction, Realtime Plus may remove the need for a separate ASR round trip.
05
Model naming and availability caveats
Launch coverage also references ASR-Next, but the callable model IDs verified here are the documented ASR-Flash variants. Do not invent an ASR-Next endpoint or substitute an inferred model ID into production code.
Region availability and protocol support can change. Confirm the exact model, region, API key scope and current API reference before deployment, especially when moving between Singapore, Beijing and other Model Studio regions.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.