Get a more accurate transcript
Speech recognition is very good at a clear voice and very bad at a muffled one. Nothing you do after the recording helps as much as the four things below do during it.
Record for the transcript
- Get the microphone close. One metre instead of three is the single biggest improvement available.
- Use Meetings & lectures mode in a room — noise suppression and echo cancellation both help recognition.
- Ask people not to talk over each other. Overlapping speech is where recognition fails completely.
- Kill steady noise: fans, air conditioning, a car engine, a café's music.
Set the right language
Recognition is language-specific. A recording in Vietnamese transcribed as English comes back as nonsense. Set Settings › Transcription language to what is actually being spoken, and transcribe a mixed-language meeting in the language most of it is in.
Help it along
- Trim the noisy lead-in before transcribing — it also saves your free allowance.
- Run noise reduction first on a recording with constant hum.
- Transcribe a long recording in parts if one section is much noisier than the rest.
- Edit the text afterwards. Twenty corrections to an hour of speech is a good trade for typing none of it.
FAQ
- Why does it get names wrong?
- Recognition works from a general vocabulary, and personal and product names are not in it. They are the first thing to fix by hand.
- Does a better recording quality setting help?
- Up to a point. 22.05 kHz is plenty for speech — position and noise matter far more than bitrate.
- Can it handle two languages in one recording?
- Not well. Recognition runs in one language at a time; transcribe twice if a recording is genuinely split between two.
Record it properly next time
Voice Recorder — Novaz records meetings, lectures and ideas in M4A or WAV, marks the moments that matter, trims and merges, and turns speech into searchable text — all on your iPhone.
Download on the App Store