Detect and Manage Speakers

Speaker detection requires Sequence 1.4.0 or later. It identifies which voice is speaking and adds speaker labels to an existing transcript. It runs on your Mac using the separate Nemotron 3 model; it is not required for transcription or text-based editing.

Install the optional model

1. Activate Sequence and connect to the internet.
2. Open Settings → Transcription → Speaker Detection.
3. Review the download size on Download Nemotron 3, then start the download.
4. Wait for downloading and verification to finish.

You can also use the speaker-model download offered in an empty transcript pane once the transcription model is installed. The models are separate downloads. Downloading a speaker model does not replace Parakeet or transcribe existing media.

Installing the speaker model turns on Automatically Detect Speakers and Remember Speaker Voices. Review both settings after installation and turn off either one if you do not want that behaviour.

Detect automatically after transcription

Leave Automatically Detect Speakers enabled in Settings → Transcription → Speaker Detection. When a new transcription completes and the speaker model is installed, Sequence runs speaker detection as a separate pass.

The transcript is available before that pass finishes. Speaker detection progress appears in Action Center. Turning this setting on does not retroactively analyse every saved transcript; use manual detection for an existing recording.

Detect speakers manually

1. Open a variant or programme with a saved transcript and accessible source audio.
2. Open the transcript actions menu and choose Detect Speakers.
3. Review the resulting labels. To run another analysis, choose Detect Speakers Again.

Manual detection still works with Automatically Detect Speakers turned off. It does not require deleting or retranscribing the words. Detection is unavailable while the current transcript is transcribing, while a detection pass is already running, or when the transcript is read-only.

Detection belongs to the shared source transcript. Variants of an ordinary source share its detected speaker facts; cameras and programmes using a Multi-Cam folder’s transcript share that analysis too. It is not a selection-only analysis of a few highlighted words.

Name and identify speakers

Open the library’s Speakers section to see recordings grouped by detected or identified person.

• Rename: right-click a speaker heading, choose Rename, enter a name, and save. A person’s name is shared across recordings linked to that identity.
• Colour and image: use Set Color, Add Speaker Image…, or the replacement/removal image actions in the same context menu.
• Assign a detected voice to an existing person: right-click the detected speaker, choose Merge with Person…, and select the intended person. This combines those identities rather than just giving them matching names.
• Correct a mistaken link: use Separate Speaker… and choose the affected recording. The separated speaker can then be named or merged with the correct person.
• Review appearances: choose Show All Projects to search for that person’s recordings.

Use the transcript filter menu's speaker entries to highlight that person's words for review. Filtering does not remove anyone from playback or export. Identity merge/separate controls change the person linked to a detected voice; they are separate from correcting a word's speaker label.

Correct who said a passage

Select a word or passage in the transcript, right-click, and open Speaker. Choose one of the transcript's existing speakers, or None to clear attribution for the selected words. A checkmark means all selected words already have that assignment.

This is an undoable transcript edit. It does not create another voice, change the recorded audio, merge people across the library, or run speaker detection again. The menu needs actual transcript words and is unavailable in read-only mode.

Remember voices locally

With Remember Speaker Voices enabled, Sequence stores local voice profiles so later detection can recognize recurring speakers. A match is a suggestion to review, not a guarantee of identity.

Voice profiles remain on this Mac even after their projects are deleted. Turning voice memory off stops learning new profiles and matching new detections against saved voices; existing person assignments remain. Use Forget Person in the library’s speaker context menu to remove a person and their learned voice data. This is different from deleting the speaker model.

An encrypted Sequence Package can carry voice-matching information for the included people when voice memory is enabled. This is an explicit project handoff, separate from cloud transcription or analytics. Unencrypted packages omit that voice evidence. The recipient can use it to match existing local people when their own voice memory is enabled; the imported evidence is not enrolled as a new local voice profile.

Limits and recovery

Speaker detection adds labels, not separate audio tracks or isolated voices. Review the result, especially around short utterances or overlapping speech. Ambiguous words can remain without a speaker assignment. Naming a speaker does not correct the recognized words; use transcript editing for text corrections.

• Model missing: install the speaker model in Settings before choosing Detect Speakers.
• Download unavailable or failed: follow the message in Settings, sign in again if requested, and retry after checking your connection. Only one model download can run at a time.
• Missing or unreadable source audio: relink the recording or select a readable audio source, then retry.
• Another model task is running: let it finish, then retry detection. Action Center shows detection failures and their available details.

To reclaim disk space, use the trash button beside the speaker model path in Settings → Transcription. Removing this download does not remove the separate transcription model; reinstall it when you want to run detection again.