Qwen Alibaba Cloud : Qwen3 ASR Flash Realtime
Qwen streaming speech recognition model for live audio over an OpenAI-realtime-style WebSocket, returning incremental transcripts with a detected language and an always-on emotion label.
Modèles variantes
Il s’agit d’ID de modèles natifs acceptés par l’auteur ou le fournisseur pour la même famille de modèles canonique.
Stable id, supplied as the model query parameter on the realtime WebSocket URL and currently equivalent to the qwen3-asr-flash-realtime-2025-10-27 snapshot.
Latest dated snapshot of Qwen3 ASR Flash Realtime, listed separately from the stable id.
Earlier dated snapshot of Qwen3 ASR Flash Realtime; the stable qwen3-asr-flash-realtime id currently resolves to this snapshot.
Fournisseurs
Plusieurs fournisseurs peuvent héberger le même modèle. Choisissez-en un lorsque vous avez besoin d’un backend fixe, ou laissez le routage choisir parmi eux.
Langues prises en charge
zh
yue
en
ja
de
ko
ru
fr
pt
ar
it
es
hi
id
th
tr
uk
vi
cs
da
fil
fi
is
ms
no
pl
sv
Fonctionnalités
Transcribes spoken audio to text.
speech_to_text
Realtime recognition over an OpenAI-realtime-style WebSocket, one session per connection, with the client appending base64 audio and the server returning incremental transcription events per turn.
streaming
Partial results arrive as transcription text events carrying a confirmed `text` prefix and a provisional `stash` suffix that concatenate to the sentence preview.
interim_transcripts
Server voice activity detection segments the stream into turns, or turn_detection null selects manual turns closed with input_audio_buffer.commit.
turn_detection
The language field is optional, and every transcription event reports the detected language alongside the text.
language_identification
Supports the 27 language codes the realtime client-events reference enumerates, with Cantonese as the separate yue code.
multilingual_transcription
Accepts free-form context text in session.input_audio_transcription.corpus.text, up to 10,000 tokens, that biases recognition toward domain-specific terms.
context_biasing
Returns an always-on emotion label (surprised, neutral, happy, sad, disgusted, angry, fearful) on every transcription event.
emotion_recognition
Code
STT en streaming avec l’ElevenLabs SDK
Diffuse l’audio du micro ou d’un fichier via la route AllModels compatible ElevenLabs.
import { ElevenLabsClient } from "elevenlabs";
const client = new ElevenLabsClient({
apiKey: process.env.ALLMODELS_API_KEY,
baseUrl: "https://api.allmodels.io/el"
});
const connection = await client.speechToText.realtime.connect({
modelId: "qwen/qwen3-asr-flash-realtime",
languageCode: "en"
}, {
queryParams: {
provider: {
"only": [
"qwen"
]
}
}
});
// Send audio frames to connection, then handle transcript events.