Speech-to-text glossary
The terms you meet when comparing speech-to-text APIs, defined precisely, especially the ones about telling speakers apart.
Speaker diarization
Working out who spoke when in a recording: splitting the audio into speaker turns and giving each turn a consistent label such as SPEAKER_1. It does not identify people by name; it tells voices apart. Transcripts of meetings, interviews, calls and podcasts depend on it. Best speaker diarization API.
Diarization error rate (DER)
The standard measure of diarization quality: the share of speaking time that is missed, falsely detected, or attributed to the wrong speaker. Lower is better. A DER figure is only comparable with another measured on the same data with the same collar. Why two providers can both claim the lowest DER.
Collar
A tolerance window, usually 0.25 seconds either side of each speaker change, that is excluded when DER is scored. It forgives disagreement about the exact moment a turn ends. A strict score uses no collar and is always higher.
Missed speech
Time where someone was talking but the system detected no speaker. One of the three parts of DER, and the one that most often drops short replies such as "yes" or "same here".
False alarm
Time the system labelled as speech when nobody was talking: music, laughter, background noise. The second part of DER.
Speaker confusion
Speech that was detected but given to the wrong speaker. The third part of DER, and the one a reader notices as a quote credited to the wrong person.
Overlapping speech
Two or more people talking at once. Overlap-aware scoring counts every speaker in the overlap, so missing one of them is an error. Dinner-party and meeting recordings have the most of it.
Word error rate (WER)
The share of words a transcript gets wrong, counting substitutions, deletions and insertions against a reference transcript. It measures the words only: a transcript can have a low WER and still credit every sentence to the wrong speaker. Benchmarks.
cpWER
Concatenated minimum-permutation word error rate: WER computed per speaker after matching hypothesis speakers to reference speakers. It mixes word accuracy and speaker attribution into one number, which makes it hard to tell which of the two went wrong.
Word-level timestamps
A start and end time for every word in the transcript. They power clip cutting, karaoke-style highlighting, search that jumps to the moment, and caption timing. Accuracy is usually reported as the median error of word boundaries in milliseconds. Subtitles and captions.
Utterance
One contiguous turn by one speaker. Speech-to-text APIs return utterances so a transcript can be rendered as a conversation without regrouping words yourself.
Batch transcription
Transcribing a finished recording, as opposed to live streaming audio as it is spoken. Batch can look at the whole file before deciding who said what, which is why it gives better speaker labels than streaming. Batch vs real-time transcription.
Code-switching
Speakers changing language mid-conversation, or mid-sentence. An API that detects one language per file transcribes the other language as nonsense; per-word language detection handles it.
SRT and VTT
The two common subtitle formats. Both are text files of numbered, timed cues; VTT (WebVTT) is the web standard used by the HTML video element, SRT the older format most video editors import. Subtitle generation API.
Custom vocabulary
A list of names and terms passed with a transcription request so they are spelled the way you spell them: product names, people, jargon. Also called keyword boosting or key-term prompting.