Qwen Alibaba Cloud: Qwen3 ASR Flash Filetrans
Qwen long-file speech recognition model served as an asynchronous transcription task, accepting files up to 2 GB and 12 hours and returning sentence-level timestamps plus optional word-level timestamps.
バリアントモデル
同じ正規モデルファミリーに対して、作成者やプロバイダーが受け付けるネイティブのモデルIDです。
Stable id, sent as the model field of the asynchronous transcription task body and currently equivalent to the qwen3-asr-flash-filetrans-2025-11-17 snapshot.
The only dated snapshot published for this model, so stable and snapshot do not diverge today.
プロバイダー
同じモデルを複数のプロバイダーがホストしている場合があります。バックエンドを固定したい場合は1つのプロバイダーを選び、そうでなければルーティングに選択を任せられます。
対応言語
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
機能
Transcribes spoken audio to text.
speech_to_text
The caller submits a file URL and receives a task id, polls the task until it reports SUCCEEDED and then fetches the result JSON, for files up to 2 GB and 12 hours.
batch
parameters.enable_words adds a words array of begin_time, end_time, text and punctuation to every sentence for ten of the 27 input languages, and also switches segmentation from VAD alone to VAD plus punctuation.
word_timestamps
Every sentence in the result carries sentence_id, begin_time and end_time in milliseconds with no flag required.
sentence_timestamps
parameters.channel_id selects which channels of a multi-channel recording to transcribe, each returned as its own transcripts[] entry and each billed separately.
multi_channel_transcription
Accepts parameters.corpus as an object of the shape {"text": "<free-form context>"}, not a bare string.
context_biasing
Optional inverse text normalization via parameters.enable_itn, off by default and applying to Chinese and English audio only.
inverse_text_normalization
parameters.language is optional, and the detected language is reported per sentence on transcripts[].sentences[].language.
language_identification
Multilingual across the 27-code Qwen-ASR language enum.
multilingual_transcription
Returns an always-on emotion label per sentence on transcripts[].sentences[].emotion, with seven values (surprised, neutral, happy, sad, disgusted, angry, fearful).
emotion_recognition
The result JSON carries a top-level audio_info object of format and sample_rate only, with no duration field.
audio_info_metadata
コード
ElevenLabs SDKでのストリーミングSTT
マイクまたはファイルの音声を、ElevenLabs互換のAllModelsルート経由でストリーミングします。
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-filetrans",
languageCode: "en"
}, {
queryParams: {
provider: {
"only": [
"qwen"
]
}
}
});
// Send audio frames to connection, then handle transcript events.