Audio transcription

Whisper-class speech-to-text. Multipart audio upload, plain-text or SRT/VTT subtitle output.

POST /api/v1/audio/transcriptions

curl https://api.pendra.ai/api/v1/audio/transcriptions \
  -H "Authorization: Bearer pdr_sk_..." \
  -F file=@meeting.mp3 \
  -F model=whisper-large-v3-turbo \
  -F language=en

from pendra import Pendra

client = Pendra()

with open("meeting.mp3", "rb") as f:
    result = client.audio.transcriptions.create(
        model="whisper-large-v3-turbo",
        file=f,
        language="en",
    )
print(result.text)

import Pendra from 'pendra';
import { readFileSync } from 'node:fs';

const client = new Pendra();

const result = await client.audio.transcriptions.create({
  file: { filename: 'meeting.mp3', content: readFileSync('meeting.mp3') },
  model: 'whisper-large-v3-turbo',
  language: 'en',
});
console.log(result.text);

200

{
  "text": "We're starting the meeting now."
}

New to transcription? Start with the Audio transcription guide. This page is the field-by-field reference.

POST /api/v1/audio/transcriptions is mirrored at the OpenAI-style alias /v1/audio/transcriptions. Multipart form upload, up to 25 MB per file; a larger upload is rejected with 413 before it is processed. Uploads must be wav, mp3, or flac — convert other formats first (e.g. ffmpeg -i input.m4a output.wav). The file's contents decide, not its name: any other format, including a browser recording saved with an .mp3 name, is rejected with 400 and the message Unsupported audio format. Accepted formats: wav, mp3, flac.

Body multipart/form-data

file file Required

The audio file (wav, mp3, or flac).

model string Required

Whisper-class model ID, e.g. whisper-large-v3-turbo.

language string

Optional ISO 639-1 code (e.g. en, cy) to improve accuracy.

prompt string

Optional priming text (acronyms, names) to bias the transcription.

temperature number default: 0

Sampling temperature. Leave it at 0 for the most accurate, fully deterministic transcript. Raising it makes the decoder sample instead of searching, which also switches beam_size off — use it only to shake a stuck or repeating transcript loose.

beam_size integer default: 5

How many candidate transcripts the decoder keeps in play. Higher is more accurate and slower; 1 decodes greedily, which is the fastest option. Range 1–8. Leave it out to get the accuracy-first default. Applies when temperature is 0 (the default).

response_format string default: json

json, text, srt, vtt, or verbose_json (timing metadata + optional word/segment timestamps).

timestamp_granularities string[]

["segment"] and/or ["word"]. Only applies with response_format=verbose_json.

Response

The shape depends on response_format. The default json returns { "text": ... } (see example). text returns a plain UTF-8 string. verbose_json adds timing and language metadata plus optional timestamps:

{
  "task": "transcribe",
  "language": "en",
  "duration": 12.48,
  "text": "We're starting the meeting now. First item on the agenda is the Q3 forecast.",
  "segments": [
    { "id": 0, "start": 0.0, "end": 4.2, "text": "We're starting the meeting now.", "avg_logprob": -0.21, "no_speech_prob": 0.01 },
    { "id": 1, "start": 4.2, "end": 12.48, "text": "First item on the agenda is the Q3 forecast.", "avg_logprob": -0.18, "no_speech_prob": 0.02 }
  ]
}

Every transcription also names the worker that produced it: in the X-Worker-Id and X-Worker-Name headers, and for the JSON formats in a pendra field too ("pendra": {"worker": {"id": ..., "name": ..., "version": ...}, "request_id": ...}). In the SDKs it's result.pendra.worker. See Which worker served the request.

Subtitle output

Set response_format=srt (or vtt) for time-coded captions. The body is the raw subtitle file:

1
00:00:00,000 --> 00:00:04,200
We're starting the meeting now.

2
00:00:04,200 --> 00:00:12,480
First item on the agenda is the Q3 forecast.

curl https://api.pendra.ai/api/v1/audio/transcriptions \
  -H "Authorization: Bearer pdr_sk_..." \
  -F file=@meeting.mp3 \
  -F model=whisper-large-v3-turbo \
  -F response_format=srt

with open("meeting.mp3", "rb") as f:
    srt = client.audio.transcriptions.create(
        model="whisper-large-v3-turbo",
        file=f,
        response_format="srt",
    )
print(srt)

const srt = await client.audio.transcriptions.create({
  file: { filename: 'meeting.mp3', content: readFileSync('meeting.mp3') },
  model: 'whisper-large-v3-turbo',
  response_format: 'srt',
});
console.log(srt);

Models

Transcription is built into every Pendra worker — no separate service to install. Two Whisper-class models are available: whisper-large-v3-turbo (fast, multilingual) and whisper-large-v3 (highest accuracy).