Skip to content

Qwen Audio 3.0 Realtime

The default voice frontend: DashScope's Qwen Audio 3.0 Realtime family, purpose-built for speech-to-speech conversation. This is what you get when you install qwen-audio-agent and only set an API key.

Models

ModelNotes
qwen-audio-3.0-realtime-plusDefault. Higher quality
qwen-audio-3.0-realtime-flashLower latency, lower cost

Both accept text and audio input, produce text and audio output, and support Function Calling (which is how the Gateway's frontend tools — task delegation, memory, reminders — reach the model).

Setup

dotenv
DASHSCOPE_API_KEY=your-key
QWEN_AUDIO_REALTIME_MODEL=qwen-audio-3.0-realtime-plus

DASHSCOPE_API_KEY can be replaced by the higher-priority alias QWEN_AUDIO_REALTIME_API_KEY when the realtime frontend needs its own credential.

One Gateway owns one active model. Switch it from the Desktop settings page or from the CLI, then restart the Gateway:

bash
qwenaudio config set --realtime-model qwen-audio-3.0-realtime-flash
qwenaudio gateway restart

WebUI and TUI only display the active model; they never override it.

Voice and turn detection

  • Default voice: longanqian — override with QWEN_AUDIO_REALTIME_VOICE.
  • Turn detection: smart_turn (semantic end-of-turn), configured by the runtime; no manual VAD tuning is exposed.

Endpoint overrides

For private deployments or proxies:

SettingEffect
QWEN_AUDIO_REALTIME_BASE_URL / QWEN_AUDIO_REALTIME_URLReplace the DashScope Realtime endpoint
DASHSCOPE_WORKSPACE_IDSwitch to a Bailian dedicated-workspace endpoint (wss://<workspace-id>.cn-beijing.maas.aliyuncs.com/...)

The transport runs 16 kHz PCM input and 24 kHz PCM output over a single WebSocket.

Capability boundary

ModelCurrent client transport
Inputtext, audiotext, audio
Outputtext, audiotext, audio

Model capability and implemented transport are deliberately tracked separately; this family has no image modality to gap. For image input at the model level, see Qwen Omni Realtime.