Glossary
What is diarization?
By the KeyMo team · Updated
Diarization is the step in speech-to-text that works out who spoke when. It splits a recording into turns and tags each one with a speaker label, such as speaker 1 or speaker 2. It tells voices apart but does not know names, so naming the people is a separate step.
What the output looks like
Google's Cloud Speech-to-Text docs describe speaker diarization as detecting when speakers change and labelling by number the individual voices it finds. Each word in the result is tagged with a speaker number, and words from the same speaker share the same number.
| Service (docs read 2026-10-11) | What you get | What you set up |
|---|---|---|
| Google Cloud Speech-to-Text | A speaker number on every word | A minimum and maximum speaker count |
| Amazon Transcribe | Labels spk_0 to spk_29 with start and end times, for up to 30 unique speakers | Turn on speaker partitioning |
The label is a tag, not an identity. Speaker 0 in one file is not the same tag as speaker 0 in the next episode.
Why it matters for two-person interviews
A transcript with no speaker labels is one wall of text. With labels you can see where a question ends and the answer starts. That decides where a podcast clip should open: on the guest's answer, or two seconds earlier on the question that makes it make sense.
Example: the host asks a 12-second question, the guest answers for 40 seconds, then the host says "right" over the end. With labels, the cut can start at the guest's first word and end before the overlap. Without them, a tool or an editor has to guess from the words alone.
- Caption colours or styles per speaker depend on the labels.
- A two-shot or multicam edit needs to know who is talking to pick the shot.
- Quotes need the right name next to them. Read speaker labels.
Where diarization goes wrong
- Crosstalk. When two people speak at once, words get assigned to one of them or to neither.
- Similar voices, such as two hosts of the same age and accent.
- Short interjections like "yeah" or "right", which carry little voice signal.
- A wrong speaker count. Google's docs make you set minimum and maximum speakers, so a bad guess can merge or split people.
- One microphone for two people in the same room, or a phone-quality remote guest.
- Languages. Google tells you to check its language support page to see if the feature exists for your language.
Not the same as face tracking
Diarization works on voices. Face tracking works on the picture. A tool can know who is speaking from the audio and still crop to the wrong face, or the reverse. KeyMo reframes to vertical 9:16 by tracking the face, so on a two-person episode it is worth checking that each clip lands on the person who is talking.
Sources: Google Cloud: Detect different speakers in an audio recording (checked 11 October 2026); Amazon Transcribe: Partitioning speakers (diarization) (checked 11 October 2026).