Click any paragraph to edit it. The title field doubles as a picker — type or select an existing transcript's title to load it. Edits save automatically — .srt/.vtt exports stay tied to the original transcription.

About History

Transcripts save automatically in this browser, newest first. Only the most recent ones (per "Keep last") are kept — older ones drop off, unless "Unlimited" is checked.

Unlimited is bounded only by this browser's local storage — commonly ~5MB per site, many hundreds of hours of transcripts in practice, but not infinite.

"Keep audio for" is separate — it's the last few *recordings* (raw audio, not just text), stored in this browser so you can re-transcribe them later with a different model. Raw audio is much larger than text, so this defaults to just 3.

Check two or more below to merge them, oldest-first, into a new transcript — the originals are removed once merged, and any audio they had moves to the merged transcript.

Deleting a transcript also deletes its stored audio. To reclaim space but keep the text, use "Delete audio" on the transcript instead. "Orphaned" recordings are audio whose transcript is already gone — older versions left these behind, and the button beside the storage total clears them.

Got it

Paste an existing transcript, or upload a .txt/.srt/.vtt file — .srt/.vtt keep their real timing, plain text gets grouped by sentence instead.

How it works

Pick "Monitor of ..." as the capture device to transcribe whatever's already playing on this machine — it's the system's own speaker output routed back in as an input, not a physical microphone. A regular microphone works too, e.g. for dictating notes directly.

System audio (a "Monitor of ..." source)
        |
        v
Web Audio API - an AudioWorklet resamples it to 16kHz mono PCM
        |
        v
A Web Worker - transformers.js loads Whisper and runs inference there,
                off the main thread, so the UI never freezes
        |
        v
Word-level timestamps -> sentences -> paragraphs (grouped by pause length)
        |
        v
Transcript UI - editable, exports to .txt / .srt / .vtt, saved locally

1. Capture system audio

const stream = await navigator.mediaDevices.getUserMedia({
  audio: { deviceId: { exact: monitorDeviceId } }
});
const ctx = new AudioContext({ sampleRate: 16000 }); // browser resamples for you
await ctx.audioWorklet.addModule('pcm-processor.js');
const source = ctx.createMediaStreamSource(stream);
const worklet = new AudioWorkletNode(ctx, 'pcm-capture-processor');
source.connect(worklet); // not connected to destination -- we're reading, not playing

2. Transcribe with Whisper, off the main thread

// whisper-worker.js -- runs inside a Web Worker
import { pipeline } from 'https://cdn.jsdelivr.net/npm/@xenova/transformers@2.17.2';

const transcribe = await pipeline('automatic-speech-recognition', 'Xenova/whisper-base.en');

const result = await transcribe(pcmFloat32Array, {
  return_timestamps: 'word',
  chunk_length_s: 30,
});

3. Turn words into readable paragraphs

function sentencesToParagraphs(sentences, pauseThreshold = 1.2) {
  const paragraphs = [];
  let current = [];
  for (const s of sentences) {
    const prev = current[current.length - 1];
    if (prev && s.start - prev.end > pauseThreshold) {
      paragraphs.push(current);
      current = [];
    }
    current.push(s);
  }
  if (current.length) paragraphs.push(current);
  return paragraphs;
}

Credits

Transcription is powered by OpenAI's Whisper speech-recognition model, running via ONNX builds published by Xenova on Hugging Face, executed entirely in-browser using transformers.js.