OpenAI Speech Recognition API (/v1/audio/transcriptions)

Transcribe audio to text, fully compatible with OpenAI's /v1/audio/transcriptions. Any OpenAI SDK can directly use it by pointing the base_url to https://api.acedata.cloud and replacing the key with your AceData Token. Supports standard complete responses and also supports gpt-transcribe's SSE incremental transcription.

• Request URL: POST https://api.acedata.cloud/v1/audio/transcriptions (alias POST /openai/audio/transcriptions)
• Authentication: Request header Authorization: Bearer {token}
• Request format: multipart/form-data
• Billing: Charged by audio duration (see table below), rounded up to the nearest second.

¶ Request Parameters

Field	Type	Required	Description
file	file	Yes	The audio file to be transcribed, maximum 25 MB. Supports flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, webm.
model	string	No	whisper-1 (default) or gpt-transcribe, see the table below for capability differences.
language	string	No	Audio language, ISO-639-1 code (e.g., zh, en). Specifying can improve accuracy and speed; leave blank for automatic detection.
prompt	string	No	Prompt words to guide writing style or provide proper nouns and terms to enhance recognition accuracy.
response_format	string	No	whisper-1: json (default), text, srt, verbose_json, vtt; gpt-transcribe: only json, text.
temperature	number	No	Sampling temperature 0–1, default 0.
timestamp_granularities[]	array	No	Timestamp granularity, word or segment, must be used with response_format=verbose_json.
languages[]	array	No	Candidate languages (ISO-639-1), only for gpt-transcribe. Mutually exclusive with language, do not send both.
keywords[]	array	No	Proper noun/term hints, only for gpt-transcribe, can significantly improve recognition accuracy for brand names and personal names.
stream	boolean	No	When set to true for gpt-transcribe, returns SSE incremental events; whisper-1 will ignore this parameter and return complete results (consistent with OpenAI's official behavior).

¶ Which Model to Choose

	whisper-1	gpt-transcribe
Price	$0.0078 / minute	$0.0059 / minute (cheaper)
Recognition Accuracy	Good	Better, especially for brand names and proper nouns
Subtitle Output (srt/vtt)	✅	❌
Word-level Timestamps	✅	❌
languages[] / keywords[]	❌	✅
Returns Detected Language	Requires verbose_json	Returns by default
SSE Incremental Return	❌ (ignored stream)	✅

Need subtitles or word-level timestamps → whisper-1; for other scenarios, recommend gpt-transcribe (more accurate and cheaper).

¶ Example

curl -X POST 'https://api.acedata.cloud/v1/audio/transcriptions' \
  -H 'authorization: Bearer {token}' \
  -F file=@audio.mp3 \
  -F model=whisper-1

Response:

{
  "text": "Ace Data Cloud Platform is testing the speech recognition endpoint. The quick brown fox jumps over the lazy dog."
}

Chinese audio is also supported without specifying the language:

{
  "text": "欢迎使用 AceData Cloud 平台,我们正在测试语音识别接口,今天是 7 月 31 号。"
}

¶ Generate Subtitles

Set response_format to srt or vtt to directly obtain a usable subtitle file:

curl -X POST 'https://api.acedata.cloud/v1/audio/transcriptions' \
  -H 'authorization: Bearer {token}' \
  -F file=@audio.mp3 \
  -F model=whisper-1 \
  -F response_format=srt \
  -o subtitle.srt

Returned content (Content-Type: text/plain):

1
00:00:00,000 --> 00:00:03,800
Ace Data Cloud Platform is testing the speech recognition endpoint.

2
00:00:03,800 --> 00:00:06,280
The quick brown fox jumps over the lazy dog.

¶ Word-level Timestamps

To get the start and end times for each word, use verbose_json with timestamp_granularities[]=word:

curl -X POST 'https://api.acedata.cloud/v1/audio/transcriptions' \
  -H 'authorization: Bearer {token}' \
  -F file=@audio.mp3 \
  -F model=whisper-1 \
  -F response_format=verbose_json \
  -F 'timestamp_granularities[]=word'

Response:

{
  "task": "transcribe",
  "language": "english",
  "duration": 6.29,
  "text": "Ace Data Cloud Platform is testing the speech recognition endpoint. The quick brown fox jumps over the lazy dog.",
  "words": [
    { "word": "Ace", "start": 0.0, "end": 0.32 },
    { "word": "Data", "start": 0.32, "end": 0.54 },
    { "word": "Cloud", "start": 0.54, "end": 0.86 }
  ]
}

¶ Streaming Transcription

gpt-transcribe can return Content-Type: text/event-stream by using stream=true. The service will send OpenAI-compatible events as is: transcript.text.delta carries incremental text, transcript.text.done carries complete text and usage indicating normal completion.

curl -N -X POST 'https://api.acedata.cloud/v1/audio/transcriptions' \
  -H 'authorization: Bearer {token}' \
  -F file=@audio.mp3 \
  -F model=gpt-transcribe \
  -F stream=true

Example of event stream:

data: {"type":"transcript.text.delta","delta":"Hello"}

data: {"type":"transcript.text.done","text":"Hello world","usage":{"type":"tokens","input_tokens":14,"output_tokens":3,"total_tokens":17}}

Receiving transcript.text.done indicates normal completion. If processing fails after the stream is established, the connection will end after the event: error event; if the client disconnects actively, this processing will be canceled and will not continue generating in the background. whisper-1 will still return a normal non-streaming response even if stream=true is passed.

¶ Using the Official SDK

from openai import OpenAI

client = OpenAI(base_url="https://api.acedata.cloud/v1", api_key="{token}")
with open("audio.mp3", "rb") as f:
    result = client.audio.transcriptions.create(model="whisper-1", file=f)
print(result.text)

¶ Price

Model	Price on this platform
whisper-1	$0.0078 / minute
gpt-transcribe	$0.0059 / minute

Charged based on the actual duration of the audio, with any duration less than 1 second counted as 1 second, and a maximum of 1 hour for a single request.

¶ Notes

• The maximum size for a single file is 25 MB. If it exceeds, please segment or compress it (usually lowering the bitrate is sufficient, as speech recognition does not have high audio quality requirements).
• gpt-transcribe supports stream=true SSE; whisper-1 will ignore stream and return the complete result.
• Parameters are consistent with OpenAI's official /v1/audio/transcriptions, and the official SDK can be used by simply changing the base_url.
• include[], chunking_strategy, known_speaker_names[], known_speaker_references[] belong to transcription models that we have not yet launched, passing them will return 400 instead of silently ignoring. Model-specific parameters (timestamp_granularities[] for whisper-1, languages[]/keywords[] for gpt-transcribe) will also return 400 when passed to unsupported models.
• Requests can be time-consuming, it is recommended to set the client timeout to no less than 300 seconds.

¶ Error Codes

Status Code	Code	Description
400	bad_request	file not provided, file cannot be parsed, or parameters are invalid (model/response_format values not supported, temperature out of 0–1, timestamp_granularities[] not paired with verbose_json, passing unsupported parameters for whisper-1).
401	authentication_failed	Invalid token.
403	used_up	Insufficient balance.
413	request_too_large	Audio file exceeds the 25 MB limit.
429	too_many_requests	Requests are too frequent, please try again later.
500	api_error	Internal server error, please try again later.