Transcribe audio

Upload an audio file and use an available Genesis speech-recognition model to return its transcript. This endpoint returns a complete synchronous response.

POST https://token.moi.matrixorigin.cn/v1/audio/transcriptions

Before you begin

  1. Configure the Genesis endpoint and credential.

  2. Select a speech-recognition model in the model marketplace and confirm that your credential can call it through List models. Preserve its public model ID.

  3. Prepare an audio file. Synchronous Qwen3-ASR Flash transcription is intended for short recordings of about five minutes or less, subject to the size limits below.

Public model IDs depend on the deployment. An upstream provider’s native model name is not a substitute.

Request parameters

Use a local WAV file and the selected public model ID:

export ASR_MODEL_ID='<公开语音识别模型 ID>'

curl --fail-with-body "$GENESIS_BASE_URL/audio/transcriptions"   -H "Authorization: Bearer $GENESIS_ACCESS_TOKEN"   -F "model=$ASR_MODEL_ID"   -F 'language=zh'   -F 'file=@sample.wav;type=audio/wav'

When using -F, let cURL generate Content-Type with its boundary. Upload one audio file per request. Do not replace the file part with JSON, a file URL, or a Base64 string.

Field

Location

Required

Description

file

Multipart file

Yes

A nonempty audio file with a correct MIME type or recognized extension.

model

Multipart form

Yes

ID, name, or alias of an enabled public ASR model available to the credential.

language

Multipart form

No

Language code, such as zh or en, using only letters, digits, underscores, or hyphens. A provider may ignore this field.

Each field may appear only once. This API is nonstreaming. Unlisted form fields are not supported options; do not add stream, response_format, or timestamp parameters.

Format

File extension

Recommended MIME type

WAV

.wav

audio/wav

MP3

.mp3

audio/mpeg

M4A

.m4a

audio/mp4

WebM

.webm

audio/webm

OGG

.ogg

audio/ogg

FLAC

.flac

audio/flac

Base64-encoded audio must not exceed 10 MiB, corresponding to about 7.5 MiB of raw audio. The deployment’s request-size limit can be smaller. Shorten or split oversized recordings rather than renaming their extensions.

Successful response

{
  "text": "你好,这是一段语音转写示例。",
  "model": "<公开语音识别模型 ID>",
  "request_id": "<上游请求 ID>",
  "usage": {"seconds": 48}
}

Field

Type

Description

text

string

Complete transcript. An upstream success without valid text is treated as an error.

model

string

Requested public model identifier.

request_id

string

Optional; returned when supplied by the provider. Preserve it for troubleshooting.

usage

object

Optional provider-reported usage. Duration in seconds can be used for per-minute billing.

Error response

Errors use the Genesis OpenAI-compatible error structure. Preserve the HTTP status, message, and available request ID.

HTTP status

What to check

400

Multipart encoding, nonempty file and model, duplicate fields, and language syntax.

401

Credential validity and a single authentication header.

404

Public model existence, enabled state, and access.

413

Encoded audio size and deployment request-size limit; shorten the input.

429

Rate-limit feedback before retrying.

501

Missing provider ASR endpoint; contact the deployment administrator.

502

Invalid upstream response or empty transcript.

503

No available provider route for the model.

A timeout does not establish that processing never ran. Inspect request records before retrying to avoid duplicate calls and charges.

Last updated on