Skip to content

1. Model Description

Muse Voice Transcribe is designed for transcribing existing recordings. For a single request, it returns the complete JSON, event stream or plain text as specified in the Accept header. It does not return a task_id, so no querying or callback is required.

ability Muse Voice Transcribe
Audio Transcription Supports uploading a WAV file via multipart and returning the transcription result synchronously
Response Pattern application/json (default or */*), text/event-stream, text/plain
Speakers and Turns ENDPOINTING: segmenting turns; DIARIZATION: adding speaker labels to turns

Supports 25 languages and allows code-switching: Arabic, Bengali, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malay, Mandarin Chinese, Marathi, Polish, Portuguese, Spanish, Tagalog, Tamil, Telugu, Thai, Turkish, and Vietnamese.

2. Billing

Muse Voice Transcribe is billed based on the processed audio, measured in whole seconds. The pricing for both streaming and non-streaming transcription is identical, with all durations rounded down to the nearest whole second.

2.1 Billing Items and Usage of the Platform

chargeItemCode Chinese name Dosage Calculation
duration (tentative, subject to change later) Not otherwise provided Calculated based on processed audio, in seconds, rounded down to the nearest integer. For JSON format, use the value of `audioDurationMs` in the response. For text/event-stream format, use the value from the last event in the stream that contains `audioProcessedMs`. The text/plain format does not have a duration field in the body, so this billing item is not applicable for this format for the time being. Failures that do not generate a transcript and 429 errors are not counted.

3. Details of Functional Interfaces

3.1 Transcribe Recordings

3.1.1 Request URL

POST https://{YOUR_DOMAIN}/v1/ai/{YOUR_ENDPOINT}/muse/v1/asr/transcribe

Header: Authorization is Bearer {YOUR_AK}. Content-Type is multipart/form-data. Accept selects the response format; a 406 status code will be returned for unsupported values.

3.1.2 Request Body Parameters / Request Query Parameters

Use multipart/form-data, which consists of two parts: the request part carries the transcription settings in JSON format, and the audio part carries the audio clip.

Parameter Name Field Type Required or not default value Description
sessionId string No - An ID used for server-side log correlation, which is returned as-is in the response. If omitted, it will be generated by the server.
request object (JSON) Yes - The request part in the multipart body is sent in JSON format.
request.audioEncoding enum ('WAV') Yes None The container for the uploaded audio part, sent as an enumeration name. WAV is a RIFF/WAVE container containing 16 kHz or 24 kHz mono integer PCM. There is no default value; values that are not set or cannot be recognized will be rejected.
request.mode enum ('PUSH_TO_TALK', 'ENDPOINTING', 'DIARIZATION') No PUSH_TO_TALK The transcription behavior is specified by an enum name. PUSH_TO_TALK corresponds to a single turn, where the entire uploaded audio constitutes one turn. ENDPOINTING detects the start and end of speech, with each detected segment treated as one turn. DIARIZATION further assigns each turn to a speaker. The available modes depend on the model. (Default: PUSH_TO_TALK)
request.keywords array of string No - Guide the model to recognize biased words, such as names, terms, and product names; otherwise, the model may mishear them. Note that biasing does not guarantee exact spelling matches.
request.languageBias array of string No - Languages to bias the transcription towards; each entry is a language name, e. g. ["English", "French"]. This is a list of languages, not free text. If omitted, the model will detect the language automatically.
request.partialMode enum ('CUMULATIVE', 'DELTA') No CUMULATIVE The representation mode of partial transcription text is sent via an enumeration name, and it only applies to text/event-stream. CUMULATIVE replaces the previous partial result; DELTA only sends the newly generated text for the caller to append. (Default: CUMULATIVE)
request.emitAudioProgress boolean No true Whether to send audioProgress events. This only applies to text/event-stream; buffered responses will not emit these events. (Default: true)
audio
binary Yes - This interface accepts RIFF/WAVE containers containing mono, 16-bit integer PCM with a sample rate of 16 kHz or 24 kHz.

Accept: application/json, */* — if this header is omitted, the complete buffered response body will be returned. text/event-stream returns a real-time event stream that terminates after the final event without appending a summary. text/plain is not currently supported.

Returns 413 if the request body exceeds 32 MB. Returns 400 if the audio exceeds 10 minutes, is not a supported WAV file, or the request body is not in multipart/form-data format.

3.1.3 Response Parameters

application/json (and */* or the default value) corresponds to TranscribeResponse.

Parameter Name Field Type Description
sessionId string Query parameter: sessionId; if the caller does not pass this parameter, it will be the ID generated by the server.
transcript string The final transcription of the entire audio, which will be populated for all modes; in multi-turn mode, it is formed by concatenating the transcriptions of each turn in sequential order.
audioDurationMs integer (int64) Total duration of processed audio, in milliseconds.
turns array of Turn All rounds reported by the model, sorted in ascending order of turnId. This field is empty under PUSH_TO_TALK. If the audio ends midway in a round, that round will carry the boundary provided by the model, with an empty transcript.
turns[].turnId integer (int32) ID of the current round. Rounds are reported in ascending order of ID; the numbering itself carries no other meaning.
turns[].startMs integer (int64) The number of milliseconds, measured from the start of the segment, of the audio that has already been processed when the model detects the start of this round of speech. This is not an exact acoustic boundary.
turns[].endMs integer (int64) The number of milliseconds, measured from the start of the segment, of the audio that has been processed when the model detects the end of this round of speech. This is not an exact acoustic boundary.
turns[].transcript string This final transcription round includes post-processing.
turns[].speaker string Speaker labels generated by the model, such as A. Only appears in DIARIZATION.

The events of text/event-stream are isomorphic to real-time server-sent events. Unknown types are ignored. The stream ends after the last event, with no appended JSON summary.

Parameter Name Field Type Description
type=transcript / transcript string Transcribed text of this segment. Under CUMULATIVE, replace the partial result of the previous entry; under DELTA, append the result.
type=transcript / final boolean Transcription stability of the model report. In PUSH_TO_TALK, true indicates the end of the stream. In ENDPOINTING and DIARIZATION, the end of a turn is determined by speechComplete.
type=transcript / audioProcessedMs integer (int64) The total duration of audio processed so far, in milliseconds.
type=speechStart / turnId integer (int32) The ID of the current speech turn. A subsequent turn may start before the speechComplete event of the previous turn is received, so the status should be maintained based on the turnId.
type=speechStart / audioProcessedMs integer (int64) The number of milliseconds, measured from the start of the stream, of the audio that had been processed when the model emitted this event; this is not an exact acoustic boundary.
type=speechEnd / turnId integer (int32) The ID of this round of voice.
type=speechEnd / audioProcessedMs integer (int64) The number of milliseconds, measured from the start of the stream, of the audio that had been processed when the model emitted this event; this is not an exact acoustic boundary.
type=speechComplete / turnId integer (int32) The ID of this round of voice. It is associated with lifecycle events in an equivalence relationship within the same session; the numbering itself carries no other meaning.
type=speechComplete / transcript string Full-turn transcription. It may differ partially from the last result, as the model may perform post-processing on this turn after the speech ends.
type=speechComplete / audioProcessedMs integer (int64) The number of milliseconds, measured from the start of the stream, of the audio that had been processed when the model emitted this event; this is not a precise acoustic boundary.
type=speaker / label string The speaker labels generated by the model, such as A, are only meaningful within the same session and do not represent verified identities.
type=speaker / audioProcessedMs integer (int64) The number of milliseconds from the start of the stream for the audio that has been processed when the model emits this tag. This is not an exact acoustic boundary.
type=audioProgress / audioProcessedMs integer (int64) The total duration of audio processed so far, in milliseconds measured from the start of the stream.
type=error / message string A human-readable error message that can be displayed to the caller.
type=error / sessionId string The session ID used for log correlation. It is empty when not yet assigned.
type=error / errorType string The type of this failure in the error classification of the model API.
type=error / errorCode string Machine-readable error code; only present when the failure is associated with an error code.
type=error / errorParam string Error-related request fields; only applicable when relevant.
Status Description
400 The audio is not a supported WAV file, the audio exceeds 10 minutes, or the request body is not multipart/form-data.
406 The Accept header has requested a media type that this interface does not produce.
413 The request body exceeds the 32 MB upper limit.
429 The caller has reached the concurrency or hourly session limit. Please retry with exponential backoff.
500 Transcription failed, or processing exceeded the server budget. Since the status code cannot be modified once the event stream has started, the failure is returned via a termination error event on an already opened stream with HTTP 200.

3.1.4 Request Example

curl --location 'https://{YOUR_DOMAIN}/v1/ai/{YOUR_ENDPOINT}/muse/v1/asr/transcribe?sessionId=my-session-id' \
--header 'Authorization: Bearer {YOUR_AK}' \
--header 'Accept: application/json' \
--form 'request={"audioEncoding":"WAV","mode":"DIARIZATION"};type=application/json' \
--form 'audio=@meeting_audio.wav'
curl --location 'https://{YOUR_DOMAIN}/v1/ai/{YOUR_ENDPOINT}/muse/v1/asr/transcribe?sessionId=my-session-id' \
--header 'Authorization: Bearer {YOUR_AK}' \
--header 'Accept: text/event-stream' \
--form 'request={"audioEncoding":"WAV","mode":"ENDPOINTING"};type=application/json' \
--form 'audio=@meeting_audio.wav'

3.1.5 Response Example

{
  "sessionId": "9f1c...",
  "transcript": "How is the weather today? It's raining.",
  "audioDurationMs": 8240,
  "turns": [
    {
      "turnId": 1,
      "startMs": 1520,
      "endMs": 4640,
      "transcript": "How is the weather today?",
      "speaker": "A"
    },
    {
      "turnId": 2,
      "startMs": 5900,
      "endMs": 8240,
      "transcript": "It's raining.",
      "speaker": "B"
    }
  ]
}
event: transcript
data: {"type":"transcript","transcript":"How is the weather","final":false,"audioProcessedMs":2400}

event: speechComplete
data: {"type":"speechComplete","turnId":1,"transcript":"How is the weather today?","audioProcessedMs":4640}
A:Hello
unknown:<transcript>