MP3 to text

How it works

1. Open this page — the tool is already set up for “MP3 to text”.
2. Drop the file in, browse for it, or paste it with Ctrl+V.
3. The model runs on your own device; the transcript appears as it goes.
4. Correct anything misheard, then download text, subtitles or notes.

What an MP3 actually gives the model

MP3 is lossy, and speech recognition cares less about that than you would think. Whisper works from a mel spectrogram sampled at 16 kHz mono — roughly the information in a telephone call plus some headroom. A 128 kbps MP3 already carries far more than that, so re-encoding a 320 kbps file down does not help and converting up from a low-bitrate original does not either. What matters is what was captured in the room, not the bitrate it was saved at.

That is why a 64 kbps MP3 of one person speaking into a headset usually transcribes better than a 320 kbps recording of four people around a table. If you have a choice at recording time, spend the effort on microphone placement rather than file settings.

Long MP3s and what the browser can hold

There is no file-size limit imposed here, but the whole recording has to be decoded into memory before transcription starts, and decoded audio is much larger than the compressed file. A one-hour MP3 is comfortable on any modern laptop. Multi-hour recordings work, but they use a lot of RAM and take proportionally longer — if a very long file makes the tab unstable, split it with any audio editor and transcribe the pieces.

The transcript arrives progressively rather than all at once, so you can start reading and correcting the beginning while the end is still running.

Getting it into a document

The plain-text export groups segments into paragraphs at natural pauses instead of emitting one line per phrase, which is the single biggest difference between a transcript you can paste into a report and one you have to reformat first. Add timecodes if you need to cite positions, or take the Markdown export straight into Obsidian or Notion.

If you are going to quote from it, use the player: click any line and the audio jumps there, so verifying a quote takes seconds rather than scrubbing blindly.

Things that save a re-run

• If the recording is stereo with one speaker per channel, that is fine — the channels are averaged, so both voices survive.
• Very quiet recordings transcribe better if you normalise the volume first in any free audio editor; the model has no automatic gain control.
• Music and long silences produce Whisper’s classic invented phrases. Repeated identical lines are filtered out automatically, but check the ends of the transcript.