Speaker labels and timestamps explained

A speaker label is a tag in a transcript that says which voice the words came from. A timestamp is the time in the recording when a word or passage starts or ends. Together they answer "who said what, and when."

What diarization means

The technical name for labelling speakers is diarization (British spelling: diarisation). Wikipedia defines it as "the process of partitioning an audio stream containing human speech into homogeneous segments according to the identity of each speaker"¹. It is "used to answer the question 'who spoke when?'"¹.

It has two parts. Wikipedia says it "is a combination of speaker segmentation and speaker clustering"¹. Segmentation finds the points where the speaker changes. Clustering groups the pieces that sound like the same person.

Labels aren't names

A tool that separates voices usually calls them by number. Amazon says its service can tell speakers apart and "labels the text from each unique speaker with a unique value"². It lists those values as spk_0 and onward, up to a stated maximum of 30 speakers on that service, as read on 2026-10-26².

The tool doesn't know that spk_1 is your sales manager. You match labels to people by listening. Wikipedia says a true identity comes from speaker recognition systems used together with diarisation, "providing the speaker’s true identity"¹.

What timestamps look like

Amazon's transcripts give a time for every word and punctuation mark³. With speaker partitioning on, the output also holds a section with "information on each utterance, including speaker label and timestamps"². An utterance is one stretch of speech by one person.

Other tools may show times only per paragraph, or not at all. Look at the maker's sample output to see what you'll get.

How reliable is it?

The sources saved for this page give no accuracy figure for labelling, so none is stated here. They do point to weak spots.

• Overlap. Google warns that speech from several people at once, or at different volumes, may be treated as background noise⁴. Dropped words can't be labelled.
• Grouping by sound. Clustering groups speech "on the basis of speaker characteristics"¹. That's a judgment made from the audio, so it can be wrong.
• Speaker count. Amazon's console asks you to specify the number of speakers when you set up speaker partitioning².

Before you rely on labels

A neat transcript with numbered speakers looks authoritative. It's still the tool's guess. Don't quote someone by name in a report because a label said so.

Spot-check a few turns against the audio, especially where someone interrupts. If a decision, quote or action item hangs on who said it, listen to that part. Pages on recording tips cover separate microphones, which help the tool.