Transcribing recordings to text — lectures, interviews, voice notes

An hour of audio takes four to six hours to transcribe by hand. Speech-recognition models do it in minutes, and the current generation (Whisper-class models) is accurate enough that you fix occasional words rather than typing everything.

The basic flow

1. Get the recording as a file — MP3, WAV, M4A, or a forwarded WhatsApp voice note all work.
2. Drop it into Audio to Text. Language is detected automatically — including Bangla, which most transcription services either skip or butcher.
3. Read through once while the audio is fresh; names and technical terms are where models guess.

What decides the quality

• Microphone distance beats everything. A phone on the table next to the speaker transcribes dramatically better than the same phone in a bag.
• Crosstalk — two people talking at once — is the hardest thing for any model. Interviews with clean turn-taking come out nearly perfect.
• Background music hurts more than background noise; it competes for the same frequencies as speech.

If you're recording specifically to transcribe, do it with the Voice Recorder — it records in-browser (nothing uploads while you record) and hands you a clean WAV.

After the transcript

The transcript is text like any other, so the rest of the toolbox applies: Summarize an hour-long meeting into decisions and action items, or Translate an interview conducted in Bangla for an English-language report — and vice versa.

Honest limits

Transcription runs on an AI provider (audio is sent for processing — not something a browser can do at this quality yet). Very poor phone-call audio, heavy dialects and overlapping speakers still need a human pass. Budget correction time at roughly 10–20% of the audio length, not zero.