Create a transcription

POST/v1/audio/transcriptions

Create a transcription.

Compatible with OpenAI's create-transcription endpoint: upload an audio file as multipart/form-data and get its text back. The provider is the one behind your credential; Sarvam and ElevenLabs answers are returned in the same shape. The file must have an audio/* content type and be at most 25 MB (1 GB on ElevenLabs), else a 400. With stream=true the response is a text/event-stream of the provider's transcription events (OpenAI only; another provider is a 501).

Authenticate with a BYOK virtual key or with x-provider and x-api-key; a PLATFORM virtual key is a 400. The key's checks, the plan's limits and the organization's approved-provider policy (403) apply. The call is priced on the audio's duration (the audio_duration_seconds you send, else the duration the provider reports) and the cost is in the x-operation-cost-usd response header. Without x-retry-count a failed call is retried up to 3 times on 429 and 5xx; a stream is not retried.

Authorization

AuthorizationBearer <slug>.<secret>
A virtual key's API key (vk-....<secret>), as Authorization: Bearer. What the OpenAI SDK sends from its api_key. Used by the inference endpoints.
orx-virtual-keyheader
A virtual key's API key in the x-virtual-key header, for clients that keep Authorization for something else.
orx-api-keyheader
Direct credentials: the provider's own API key in x-api-key, with the provider named in x-provider (for example openai). On /v1/messages a virtual key is accepted in x-api-key too, which is what the Claude SDK sends.

Headers

x-providerstring
With direct credentials: the provider x-api-key belongs to, for example openai or anthropic (case and dashes are normalised). Not used with a virtual key.
x-retry-countinteger<int32>
How many times to retry a failed provider call, 0 to 5. Without the header, chat completions, Claude Messages and prompt completions do not retry, and the other endpoints retry up to 3 times. Retries back off exponentially from 1 second and honour a numeric Retry-After. The number made is in the x-retry-count response header.
x-retry-status-codesstring
Comma-separated provider status codes to retry on. Default 429,500,502,503,504,529.

Request body multipart/form-data

filestring<binary>required
the file to upload
modelstringrequired
the model, for example whisper-1 or gpt-4o-transcribe
languagestring
the audio's language as an ISO-639-1 code, for example en; optional
promptstring
text to guide the model's style or continue a previous segment; optional
response_formatstring
json (the default) or verbose_json, which adds the language, duration, segments and words (transcriptions only)
temperaturestring
sampling temperature from 0 to 1; optional
timestamp_granularitiesstring
word or segment, with verbose_json: the timestamps to include
audio_duration_secondsstring
the audio's length in seconds, used to price the call when the provider does not report it (always on a stream); optional
streamstring
true to stream the transcription as server-sent events; default false

Responses

200application/json

The transcription, or a stream of transcription events with stream=true

textstringrequired
The transcribed text.
languagestring
The detected or given language of the audio. With verbose_json only.
durationnumber<float>
The audio's length, in seconds. With verbose_json only.
segmentsobject[]
The transcription split into timed segments. With verbose_json and segment timestamps only.
Child attributes
- idinteger<int32>
  The segment's index, from 0.
- seekinteger<int32>
  The seek offset of the segment, in the model's audio frames.
- startnumber<float>
  When the segment starts, in seconds from the beginning of the audio.
- endnumber<float>
  When the segment ends, in seconds from the beginning of the audio.
- textstring
  The segment's text.
- tokensinteger<int32>[]
  The token ids of the segment's text.
- temperaturenumber<float>
  The sampling temperature used for this segment.
- avg_logprobnumber<float>
  The average log probability of the segment's tokens. Below -1, the segment may be unreliable.
- compression_rationumber<float>
  The segment's compression ratio. Above 2.4, the text may be repetitive or unreliable.
- no_speech_probnumber<float>
  The probability, 0 to 1, that the segment contains no speech.
wordsobject[]
Every word with its timing. With verbose_json and timestamp_granularities including word only.
Child attributes
- wordstring
  The word as transcribed.
- startnumber<float>
  When the word starts, in seconds from the beginning of the audio.
- endnumber<float>
  When the word ends, in seconds from the beginning of the audio.

Request

curl -X POST "https://api.vectoraxis.ai/v1/audio/transcriptions" \
  -H "Authorization: Bearer $VECTORAXIS_VIRTUAL_KEY" \
  -F "file=@./file" \
  -F "model=string" \
  -F "language=string" \
  -F "prompt=string" \
  -F "response_format=string" \
  -F "temperature=string" \
  -F "timestamp_granularities=string" \
  -F "audio_duration_seconds=string" \
  -F "stream=string"

Response 200

{
  "text": "string",
  "language": "string",
  "duration": 0,
  "segments": [
    {
      "id": 0,
      "seek": 0,
      "start": 0,
      "end": 0,
      "text": "string",
      "tokens": [
        0
      ],
      "temperature": 0,
      "avg_logprob": 0,
      "compression_ratio": 0,
      "no_speech_prob": 0
    }
  ],
  "words": [
    {
      "word": "string",
      "start": 0,
      "end": 0
    }
  ]
}

VectorAxis is built by GK Labs, a member of NVIDIA Inception.