MultiKit

Speech to Text: Transcribe Audio in Your Browser

  • Free
  • No sign-up
  • Runs in your browser
  • Version 1.0.0
  • Last updated 2026-10-15
Report a problem

Drop an audio file, or a video (just its sound is used), and it is turned into text with a timestamp before each part. The model runs on your device, so the recording itself is never sent anywhere; only its own files are fetched once, from Hugging Face's public model store.

Everything you enter is processed in your browser and is never sent to a server.

How to use

  1. Drop an audio or video file, or click the box to choose one.
  2. Pick the spoken language if you know it, or leave Auto-detect.
  3. Press Transcribe. The first time, the speech model loads (about 40 MB); the browser keeps it, so later files start at once.
  4. Read the transcript, copy it, or download it as a .txt file.

Examples

A two-minute voice memo in English
→ A transcript with a timestamp, such as [0:00], before each part
Longer recordings are broken into chunks behind the scenes and stitched back into one transcript, so there is no real length limit beyond patience and your device's speed.
A recording in Korean, with Auto-detect left on
→ A Korean transcript
Auto-detect works well once there are a few seconds of clear speech; for a short or noisy clip, picking the language directly can help.
A silent or music-only file
→ "No speech was found in this file."
The model is built to recognize speech, not to describe other sounds, so it says plainly when it found none rather than invent a transcript.

How it works

The model is Whisper, in its smallest published size (“tiny”), run by Transformers.js, an open-source library that loads a machine-learning model and the WebAssembly program that runs it, both fetched once from the Hugging Face Hub and kept afterward by the browser’s own cache. The audio file is decoded by the browser’s own Web Audio API into the single-channel, 16,000-sample-a-second format the model expects; nothing about this step involves a server either.

Recordings longer than 30 seconds are cut into overlapping 30-second pieces behind the scenes, transcribed one at a time, and stitched back into a single transcript with one continuous set of timestamps, the same approach Whisper itself uses for long audio. This is a small model, and its transcript is a best effort: it can mishear words, especially with background noise, overlapping voices or an uncommon language, so read it over before relying on it.

Sources

Frequently asked questions

Is my recording uploaded?
No. The audio is decoded and transcribed on your own device by a model running in your browser. The only things fetched over the network are the model's own files, the first time you use the tool; your recording never leaves your device.
What does "the speech model loads" mean, and why 40 MB?
The tool uses Whisper, an open-source speech-recognition model, in a small size chosen to load reasonably fast and run in any browser. The browser's own cache keeps it after the first time, so it is not downloaded again on your next visit, even to a different tool here that happens to use it.
How accurate is it?
Good for clear speech in a fairly common language, less good with heavy accents, overlapping speakers, background noise or music, or uncommon languages. This is a small model, chosen to be fast and to run anywhere, so it will not match a large paid transcription service on hard audio. Read the transcript before relying on it.
What file types work?
Whatever your browser itself can decode the audio from: MP3, WAV, M4A, OGG and most common formats, plus the audio track of most video files. If a file does not work, the browser could not decode it, which this page tells you rather than guess at a transcript.
Does it tell different speakers apart?
No. It transcribes what is said, with timestamps for when, but does not label who is speaking.
Version history
  • 1.0.0 — First release.

Last updated: