WHISPERX GUIDE
WhisperX diarization: make speaker labels useful
A diarized transcript is easier to read when labels are checked and named. Start with clear turns, then examine the ambiguous parts of the conversation.
By the WhisperX team · Reviewed October 15, 2026
Understand what the labels mean
Diarization is the grouping of speech by voice within a recording. Speaker 1 and Speaker 2 are task specific labels. They do not tell you who a person is, whether a voice is genuine, or whether the same person appears in another file.
Let the first result guide your review
Use automatic speaker discovery when the number of voices is uncertain. If you know the participants, compare the resulting groups with a few clearly spoken turns. Short greetings or background voices may become extra groups; two similar voices may also merge.
Name the speakers after listening
Choose a long, clear turn for each participant and listen before renaming the label. Keep names consistent across the transcript. When a label is wrong for a particular turn, correct that turn without guessing the identity of an unclear voice.
Check overlapping speech explicitly
Interruptions are hard for both recognition and speaker assignment. Slow down the review around handoffs and brief replies. If the recording does not make attribution clear, keep the uncertainty visible instead of assigning a confident name.
Know what an edit changes
Renaming a speaker changes a display name. Editing a sentence changes text, and its old word timing may no longer match. Neither action trains the model or runs diarization again. Keep the original JSON if your research or application needs to distinguish model output from a corrected version.
Running diarization locally
A local installation needs the model access and credentials described in the upstream README. The hosted workflow manages that environment for you. If you install locally, confirm which speaker model your selected release uses and read its license and access conditions.
Read the upstream diarization setup ↗