How to Use Qwen Audio 3.0 Realtime Plus and Realtime Flash

Qwen Audio 3.0 Realtime Plus and Qwen Audio 3.0 Realtime Flash are available on EmpirioLabs. Both are Alibaba voice conversation models: you open a WebSocket, stream microphone audio up, and the spoken reply streams back. Each has 19 voices, turn detection that tells a pause from the end of a sentence, native web search, and function calling. Realtime Plus balances reasoning with conversational rhythm; Realtime Flash is tuned for fast replies, at a much lower price.

Try them in the Playground: Realtime Plus or Realtime Flash. Click Start session and talk.

Watch the launch film

Every voice in the film is a real Qwen Audio 3.0 Realtime reply over the socket, each answering a typed question: Plus with loongeva and, in Mandarin, longanlingxi, and Flash with daniel and sherry. The pictures are Qwen Image 2.1 outputs, and the score was generated by ACE-Step on EmpirioLabs. Watch it on YouTube.

Connecting

wss://api.empiriolabs.ai/v1/realtime?model=qwen-audio-3-0-realtime-flash
Authorization: Bearer YOUR_EMPIRIOLABS_API_KEY

Configure the session with session.update, append microphone audio as input_audio_buffer.append, and read the reply from response.audio.delta, with its transcript on response.audio_transcript.delta. Audio travels as 16-bit PCM, 16 kHz going up and 24 kHz coming back.

{"type": "session.update", "session": {
    "voice": "daniel",
    "instructions": "You are a friendly support agent for a bike shop.",
    "turn_detection": {"type": "smart_turn"},
    "speed": 1.1
}}

Voices and delivery

• Voices. 19 voices, including the English voices daniel, echo, hannah, and sherry, listed on the voice parameter in the API reference and in the Playground picker. Leave voice unset and longanqian speaks. These voices differ from the Qwen Audio 3.1 realtime models' set.
• Delivery. speed (0.5 to 2) sets the speaking rate, enable_speech_emotion the expressive delivery, and max_history_turns (1 to 50) how many earlier exchanges the model reads.
• Turn detection. server_vad, the default, ends your turn after a pause; smart_turn also judges what was said, so filler sounds do not end a turn, and smart_turn_v2 also answers bystanders less often.

Web search or function calling

Web search is off by default; turn it on with enable_search, or declare functions in tools. Use one per session: a session.update that would turn on both is refused with an error event naming the field, and the session keeps its earlier settings. To search in a session that declared tools, send "tools": [] first. For both in one session, use Qwen Audio 3.1 Realtime Flash.

Operational notes

• Choose the voice before the first turn. Once the session has started processing audio, a session.update with a different voice is refused with an error event. Open a new session to switch voices.
• There are no language settings. input_language and output_language are Qwen Audio 3.1 settings; on 3.0 each gets an error event naming it, and the rest of the update applies. Say in instructions which language to answer in.
• Keep streaming briefly after the speech ends. Turn detection ends a turn by hearing the speaker stop, so send about a second of silence after the last word.

Pricing

Each completed turn is billed from the usage the model reports, with separate rates for audio and text tokens in both directions, so a text-only reply costs less than a spoken one. Search results count as input tokens. Current rates are on the Realtime Plus and Realtime Flash model pages and the pricing page.

Start building

The API reference for each model is at Realtime Plus and Realtime Flash, and the full event tables are in the Realtime Voice API guide. For the newer generation, see How to Use the Qwen Audio 3.1 Realtime Flash API.