ToolsMaxxing
English
EnglishEspañolPortuguêsBahasa Indonesiaहिन्दीবাংলাРусскийTiếng ViệtFrançaisDeutschTürkçeالعربية

Audio to text

Turn speech in an audio file into text with timestamps, and save it as .txt, .srt or .vtt subtitles. The AI runs on your device, so the recording is never uploaded.

or drag and drop audio files here

MP3, WAV, M4A, AAC, OGG or FLAC audio · Multiple files supported

Your files stay on your device

  1. 1Add files
  2. 2Adjust
  3. 3Download

How it works

  1. 1Choose one or more audio files: MP3, WAV, M4A, OGG, FLAC and more. You can add up to 5.
  2. 2Pick the spoken language, or leave it on detect. Choose Fast for a quick draft, Accurate for fewer mistakes, or Bangla only for Bangla.
  3. 3Select Transcribe. The first time, the AI model downloads once and your browser keeps it for next time.
  4. 4Read the text, copy it, or download .txt, .srt or .vtt. Check names and numbers against the recording.

What the options do

Model
Fast is small and quick and is fine for clear English. Accurate is bigger and makes fewer mistakes with accents, noise and most languages. Best is the largest model and needs a computer with WebGPU. Bangla only is a model trained for Bangla: choose it for Bangla speech, because the other models are not good at Bangla.
Spoken language
Choose the language of the recording. Detecting it automatically works for long clear speech, but it can pick a neighbouring language for short or noisy clips, so choosing it is more reliable.
Result
Text in the spoken language writes what is said. Translate to English writes an English version of speech in any language.
Show times in the text
Adds the start time before every line of the text box and the .txt file. The .srt and .vtt subtitle files always carry times.

Privacy

This tool runs in your browser. Your files are never uploaded to a server.

Questions people ask

Is my recording uploaded?

No. The speech model runs inside your browser, on your graphics card when your browser supports WebGPU and on your processor otherwise. The recording never leaves your device. Only the model files, which hold no personal information, are downloaded once from Hugging Face and kept by your browser.

How accurate is it, and what about Bangla?

Clear English speech gives good results with every model, and Accurate also writes punctuation well. Bangla is different. The general models are weak in Bangla: Fast writes the words in Latin letters or the wrong script, and Accurate and Best often give wrong or repeated words. For Bangla choose the Bangla only model. In our tests on a human-read news item it was readable and mostly right, but it still made mistakes in names, rare words and long compound words, and sometimes garbled a few words at the end of a phrase. Treat Bangla text as a draft and read it against the recording. Background noise, music, strong accents and several people talking at once lower accuracy in every language.

Which AI model is used, and can I use the results commercially?

It uses OpenAI Whisper models, released under the MIT licence, which allows commercial use. The text and subtitles you create are yours.

Is there a time limit?

No limit of ours. The only limit is your device: a long recording takes longer and needs more memory. A one hour recording works on a computer with a graphics card. On a phone, keep recordings short.

What are the .srt and .vtt files?

They are subtitle files with a start and end time for each line. SRT works in most video editors and players. VTT is the web version used by HTML video. Both open as plain text.

Does it tell who is speaking?

No. It writes the words in order with times, but it does not label speakers.

Related tools