File formats and audio

Sample rate

How many times per second the audio signal is measured, given in hertz.

·Also called: sampling rate

The sample rate is how many times per second the audio signal is measured when it is digitised. It is given in hertz: 16,000 Hz means 16,000 measurements per second.

The common values

RateUsed for
8 kHzTelephony. Noticeably muffled
16 kHzStandard for speech recognition
44.1 kHzCD quality, music
48 kHzVideo and professional audio work

Why 16 kHz is enough for speech

A sample rate can reproduce frequencies up to half of itself. 16 kHz therefore covers up to 8 kHz, and practically all the information that distinguishes speech sounds from one another lies below that.

Most speech models, Whisper included, convert everything to 16 kHz internally anyway. Supplying audio at 48 kHz therefore does not give better transcription, only larger files.

When it is too little

8 kHz, that is, telephone quality, cuts away the top register. That is where the difference between s and f sounds lives, among others. Recordings of phone calls are therefore transcribed noticeably worse than recordings of the same conversation in a room.

It is also why a video meeting can give better audio than a phone call, even though the distance to the microphone is the same.

In practice

Do not think about this. The default settings on phones, recorders and meeting tools are well above what is needed. The only thing to avoid is picking a “save space” setting that drops the rate to telephone quality.

See also

Audio format, word error rate, voice activity detection.

Speech, written out

Sayable transcribes speech with dialect, tells the speakers apart and drafts the summary. Free to get started.