Skip to main content

Speaker Identification

When you review a meeting transcription in Copera, each spoken segment is labeled with the speaker who said it. This is powered by a combination of real-time active-speaker tracking during the meeting and AI-based diarization during post-processing.

How it works​

Speaker identification happens in two stages:

1. Real-time tracking during the meeting​

While the meeting is in progress and transcription is active, the frontend tracks which participant is speaking at any given moment using LiveKit's active-speaker detection. This data -- a timeline of who was talking when -- is sent to the backend and stored alongside the audio recording.

2. Post-processing with AI diarization​

After the transcription recording stops, the audio is sent to the speech-to-text service, which performs its own speaker diarization -- splitting the audio into segments and labeling each segment with a generic speaker label (Speaker A, Speaker B, etc.).

The system then runs a matching algorithm that compares the real-time active-speaker data from step 1 with the diarization output from step 2. By finding time overlaps, it automatically maps the generic labels to actual participant names.

Viewing speakers in the transcript​

When you open a transcript in the session viewer, the Speakers panel on the left shows all identified speakers. Each speaker is represented by their avatar (if matched to a workspace member) and is assigned a unique color. Utterances in the transcript are color-coded to match, making it easy to scan through a long transcript and see who said what.

You can click on a speaker in the Speakers panel to filter the transcript to only show utterances from that person. Click again to clear the filter and show all utterances.

Correcting speaker assignments​

Sometimes the automatic matching may not be perfectly accurate -- especially in meetings with many participants who speak in quick succession. You can manually reassign any speaker label:

  1. In the Speakers panel, click the avatar next to a speaker label.
  2. A dropdown appears showing all meeting participants.
  3. Select the correct participant, and all utterances from that speaker are updated immediately.
  4. You can also add a custom speaker (by name and email) for external participants who were not workspace members.

When you open a speaker selection dropdown, the transcript automatically filters to that speaker and plays their first utterance so you can hear their voice and confirm the correct assignment.

Tips for better accuracy​

  • Avoid crosstalk -- when multiple people talk at the same time, the system has a harder time distinguishing speakers.
  • Use individual microphones -- if participants share a single microphone (e.g., in a conference room), the system may group them as one speaker.
  • Longer recordings help -- the more audio the system has, the more accurately it can identify distinct voice patterns.