> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.empiriolabs.ai/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.empiriolabs.ai/_mcp/server. # Qwen Audio 3.0 ASR Stream > Live transcription over a socket, streaming partial words while the speaker talks and a settled sentence at each pause, priced per hour of audio. ![Qwen Audio 3.0 ASR Stream](https://media.empiriolabs.ai/model-logos/qwen.png) [Alibaba Cloud](/providers/alibaba) · Transcription `WEBSOCKET /v1/realtime` Live transcription over a socket, streaming partial words while the speaker talks and a settled sentence at each pause, priced per hour of audio. ## At a glance | Field | Value | | ------------------- | -------------------------------------------------------------------- | | Model id | `qwen-audio-3-0-asr-stream` | | Model release date | 2026-07-30 | | Input modalities | Audio | | Output modalities | Text | | Context window | - | | Weight precision | - | | Region | Singapore | | Features | transcription, speech_to_text, realtime, streaming_asr, multilingual | | Native inference | No | | New | Yes | | Supported endpoints | `WEBSOCKET /v1/realtime` | | Alternate model ids | `qwen-audio-3.0-asr-stream`, `alibaba/qwen-audio-3-0-asr-stream` | ## Pricing | Charge | Spec | Rate | | ------------- | ----------------- | ------- | | Transcription | per hour of audio | \$0.648 | ## Example request ```python import asyncio, json, os, websockets URL = "wss://api.empiriolabs.ai/v1/realtime?model=qwen-audio-3-0-asr-stream" async def main(): async with websockets.connect( URL, additional_headers=[("Authorization", f"Bearer {os.environ['EMPIRIOLABS_API_KEY']}")], max_size=None, ) as ws: print(json.loads(await ws.recv())["type"]) # session.created await ws.send(json.dumps({ "type": "conversation.item.create", "item": {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "Say hello."}]}, })) await ws.send(json.dumps({"type": "response.create"})) async for raw in ws: event = json.loads(raw) if event["type"] == "response.done": break asyncio.run(main()) ``` ## Parameters | Parameter | Type | Required | Default | Description | | --------------------- | ------- | -------- | ------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `input_audio_format` | enum | no | `"pcm16"` | Encoding of the audio you append to the input buffer: base64 16-bit PCM, mono, at 16 kHz. · Allowed: `pcm16` | | `language` | enum | no | `"auto"` | Language of the speech, or auto to detect it from the audio. Set it with the language query parameter of the connection URL, or as input_audio_transcription.language in session.update. · Allowed: `auto`, `zh`, `en`, `ja`, `ko`, `vi`, `th`, `id`, `ms`, `tl`, `hi`, `ar`, `fr`, `de`, `es`, `pt`, `ru`, `it`, `nl`, `sv`, `da`, `fi`, `no`, `el`, `pl`, `cs`, `hu`, `ro`, `bg`, `hr`, `sk` | | `prompt` | string | no | - | Text that guides recognition, up to 400 characters: the names and terms the speaker uses, or the text spoken just before this audio. Write the words exactly as they should appear in the transcript. Also accepted as input_audio_transcription.prompt. | | `silence_duration_ms` | integer | no | `800` | The pause, in milliseconds, that ends a sentence, from 200 to 6000. A shorter pause settles each sentence sooner; a longer one keeps a pause inside a sentence. Also accepted as turn_detection.silence_duration_ms. · Range: 200 – 6000 | | `threshold` | number | no | - | How clear speech must be before it is transcribed, from 0.0 to 1.0. A higher value ignores more background noise and can miss quiet speech. When omitted, the model's own default applies. Also accepted as turn_detection.threshold. · Range: 0 – 1 | | `noise_reduction` | enum | no | `"far_field"` | The microphone the speech detection is tuned for: near_field for a headset or a phone held close, far_field for a laptop or room microphone. Also accepted as input_audio_noise_reduction.type. · Allowed: `far_field`, `near_field` | | `keep_dialect` | boolean | no | false | Keeps Chinese dialect words as spoken. When off, dialect speech is written as standard Mandarin. | ## Notes **Connecting** Live transcription over a WebSocket at `wss://api.empiriolabs.ai/v1/realtime?model=qwen-audio-3-0-asr-stream`, authenticated with the ordinary Authorization Bearer header. **Sending audio** Stream 16 kHz mono pcm16 audio with `input_audio_buffer.append`. Voice activity detection decides where each sentence ends, so there is no commit to send. **Session settings** Send `session.update` right after `session.created`, before the first audio, with any of these settings: `input_audio_transcription.language` (or the `language` query parameter of the connection URL; `auto` by default), `input_audio_transcription.prompt` with the names and terms the speaker uses (up to 400 characters), `turn_detection.silence_duration_ms` (200 to 6000, default 800) and `turn_detection.threshold` (0.0 to 1.0), and `input_audio_noise_reduction` (`near_field` or `far_field`). `keep_dialect` keeps Chinese dialect words as spoken. A setting changed while audio is streaming ends the sentence in progress and applies to the audio sent after it. **What comes back** `conversation.item.input_audio_transcription.delta` events carry the sentence so far while the speaker is still talking, and `conversation.item.input_audio_transcription.completed` carries the settled sentence at each pause. **Billing** Billed per second of the audio you stream, silence included, at the hourly rate shown. The session is billed when it ends. --- *Machine-readable schema:* `GET https://api.empiriolabs.ai/v1/models/qwen-audio-3-0-asr-stream`.