Skip to content

Qwen Streaming ASR

Uses Alibaba Cloud Qwen Server VAD to detect speech boundaries while producing preview and final transcripts. It reports facts only: speech.started/stopped are observational Events and final transcripts go to builtin.voice_turn_controller for turn admission and cancellation decisions.

Property Value
Node type qwen.asr_realtime
Layer / kind algorithm / transform
Capability speech.vad_asr.streaming

Ports

Port Direction Schema
audio_in Input Audio PCM S16LE, 16 kHz, mono, streaming
speech_out Output Event Server-VAD speech.started / speech.stopped
signal_out Output Signal Legacy graphs only; new graphs leave it disconnected and legacy mode disabled
transcript_preview_out Output Text Partial transcript
text_out Output Text Final transcript committed after Server VAD closes the turn
event_out Output Event Transcript failure state

Configuration

model defaults to qwen3-asr-flash-realtime; language defaults to zh. Demo 2 uses a vad_threshold of 0.45: lower values trigger speech more easily, while higher values reject more low-energy sounds. silence_duration_ms tunes utterance completion. To tune sensitivity, select qwen-vad-asr on the Studio canvas, change vad_threshold in Configuration, then select Validate and Save graph. Configure the shared dashscope Connection. The Node emits provider-neutral semantic Frames only. Demo 2 fans its Text and Event outputs into both Graph processing Nodes and the project-local Voice Room protocol Node. If vendor events are reordered, the Node waits for speech.stopped before committing text_out and drops previews that arrive after Final, so the Graph needs no separate Turn Context Node. The ASR provider exposes no cancellation Signal and makes no filler, cough, or short-utterance decision. Admission and cancellation policy belongs exclusively in the Voice Turn Controller.

See Alibaba Cloud's Qwen real-time speech recognition guide for the current protocol and model scope.