Speech recognition in your browser
This tool converts MP3 to text using Whisper, an open speech recognition model. Most online transcription services upload your audio to a server. Here the model is downloaded to your browser once and then runs on your own computer or phone. Your recording never leaves your device, which matters for interviews, medical or legal notes, meetings and personal voice messages. After the first download, the model is stored by the browser and loads faster next time.
Languages and accuracy
Here Whisper works with 23 languages, including English, Spanish, French, German, Italian, Portuguese, Russian, Chinese, Japanese, Korean, Turkish and Polish. Accuracy is best for clear speech with little background noise. It is lower for:
- several people speaking at the same time;
- strong background music or noise;
- very quiet or distant speakers;
- uncommon names and technical terms;
- mixing several languages in one sentence.
Selecting the language yourself, instead of automatic detection, usually gives better results, especially for short recordings.
Output formats
| Format | What it contains | Use |
|---|---|---|
| TXT | plain text | notes, documents, quotes |
| SRT | numbered subtitles with start and end times | video editors, YouTube, most players |
| VTT | web subtitles with times | HTML5 video players, web pages |
Speed and long recordings
Transcription is heavy work for a computer. On a modern laptop, a one-hour recording can take from several minutes to longer than real time, depending on the processor, the model size and the browser. Phones are slower. Keep the tab open while it works. For very long files, you can also cut the recording into parts with Split Audio and transcribe them one at a time.
Tips for a better transcript
The model listens in short windows of about 30 seconds and writes what it hears with start and end times. It adds punctuation and capital letters by itself, but it does not know who is speaking, so in an interview you will need to mark the speakers yourself. Long silences and music without speech are skipped or left empty. If the beginning of a recording is music, the first lines of text may start later than you expect; that is normal.
- Clean noisy recordings with Noise Reduction before transcribing.
- If the speaker is quiet, raise the level with Normalize first.
- For subtitles, check the timing in your video player. If one subtitle is too long to read, split it in a text editor; SRT and VTT are plain text files.
- Always read the transcript once. Correct names, numbers and words that the model may have misheard.
Video files work too: drop an MP4 and the tool transcribes its sound track, which is a quick way to make subtitles for your own videos.