Audio transcription

Pendra transcribes spoken audio into text using Whisper-class models — meetings, calls, voice notes, podcasts. It's OpenAI-compatible, so existing transcription code works against Pendra by swapping the base URL and key.

Transcribe a file

Upload an audio file and a model. Files can be up to 25 MB and must be wav, mp3, or flac — convert other formats first (e.g. ffmpeg -i input.m4a output.wav). An optional language hint (ISO 639-1, e.g. en, cy) improves accuracy.

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);

Output formats

Choose the shape with response_format:

• json (default) — a simple { "text": ... } object.
• text — the raw transcript as a plain string.
• srt / vtt — time-coded subtitle files, ready to ship as captions.
• verbose_json — timing and language metadata, plus word- or segment-level timestamps when you set timestamp_granularities.

Accuracy vs. speed

By default Pendra decodes every file with beam search, which weighs several candidate transcripts against each other and picks the best one. It is the most accurate setting and it is what you get if you send nothing.

When you would rather have the transcript back sooner — bulk jobs, rough drafts, anything latency-sensitive — set beam_size=1 to decode greedily:

Beam search applies at temperature=0, the default. A higher temperature makes the decoder sample instead, so leave temperature alone unless you are trying to shake a stuck transcript loose.

result = client.audio.transcriptions.create(
    model="whisper-large-v3-turbo",
    file=f,
    beam_size=1,  # greedy — fastest; omit for the accurate default
)

const result = await client.audio.transcriptions.create({
  file: { filename: 'meeting.mp3', content: readFileSync('meeting.mp3') },
  model: 'whisper-large-v3-turbo',
  beam_size: 1, // greedy — fastest; omit for the accurate default
});

Available models

• whisper-large-v3-turbo — fast, multilingual, the balanced default (~1.6 GB).
• whisper-large-v3 — highest accuracy, larger and slower (~3.1 GB).

Transcribe live from your microphone

Prefer to talk instead of upload? The Playground in your Pendra dashboard can transcribe straight from your microphone in real time: open the Playground, pick a transcription model, and press Record — the transcript appears as you speak. It's the fastest way to try a model on your own voice before wiring up the API.

Live transcription always decodes greedily so words appear as you speak, so expect a file transcription of the same audio to be a little more accurate. When accuracy matters more than immediacy, record first and upload the file.