Speaker Recognition & Labeling — Quill Documentation | Quill
Quill separates voices in your meeting and labels each one. This is called speaker diarization.
Over time, Quill will learn the voices of people you frequently meet and recognize them in future conversations. This all happens locally on your computer and out of the cloud.
Correcting a speaker label
If Quill labels a speaker incorrectly:
- Open the meeting.
- Click "Label Speakers"
- Quill will walk you through the phrases that it was least confident in and give you the opportunity to name the speaker, map to an existing contact, or swap with another speaker.
- If Quill thinks there are more speakers than actually participated, you can click one of the speakers in the meeting heading and merge them with one of the real participants.
Assigning a segment to a contact also teaches Quill that voice. Future meetings will recognize them automatically.
Tips for better recognition
- Add the people you meet with regularly as contacts. → Contacts
- Correct mislabeled speakers in the first few meetings with a new person. After 2-3 corrections, Quill will recognize them on its own.
- Use a quality microphone. Voice matching degrades with background noise and crosstalk.
Common questions
Why is everyone labeled "Speaker X"?
Quill hasn't matched the voice to any contact yet. Assign segments to the right contact and Quill will learn.
Can I merge two speakers Quill thinks are different?
Yes. Open the speakers panel, select the duplicate, and reassign their segments to the correct speaker.
Does voice data leave my device?
No. Voice embeddings live in your local database alongside your meetings.
Quill confused two people with similar voices.
Correct a few segments manually. The model updates from your corrections.
Quill thinks it heard someone who I know I will never see again.
You can tell Quill to not auto-recognize a specific contact in the future. View the contact > edit > uncheck "Auto-recognize this speaker in future recordings".