Shuole

Speaker Diarization

Identify and label different speakers in your audio — who said what.

What It Does

The Diarize stage analyzes your audio to detect different speakers and labels them as "Speaker 1", "Speaker 2", etc. This enables speaker-aware features throughout the app:

  • Color-coded speaker segments in the transcript
  • Speaker View mode for easy reading
  • Filter and play only specific speakers
  • Export with speaker labels included

Working with Speakers

Once diarization is complete, you can use these features on the Results page:

Speaker Mapping

Replace generic labels like "Speaker 1" with real names from your speaker database. See the Quick Start guide for details.

Speaker Filtering

Click on a speaker label to highlight and play only that speaker's segments. The audio player automatically seeks and skips to relevant parts. See View Modes in Quick Start.

Audio Clip Extraction

Extract audio clips from specific speaker segments to build your speaker database. This is useful for creating voice samples for future reference.

Audio clip extraction

Speaker Database

The Speakers page lets you manage a database of known speakers. You can:

  • Add speakers with names
  • Upload or extract audio clips for each speaker
Speaker database management

Limitations

Important: Speaker diarization is not perfect. You may encounter these issues:

  • Different people may be recognized as the same speaker (merged)
  • One person may be split into multiple speaker labels (fragmented)
  • Overlapping speech can cause misattribution

Diarization works best in settings with fewer than 15 speakers and clear turn-taking (minimal overlapping speech).

Future Improvements

We plan to add editing tools that let you manually split and merge speaker tracks to fix diarization errors easily. Stay tuned for updates.

Configuration Options

Default settings work for most cases. Click the Advanced button to reveal additional options.

Diarization configuration options

Default Options

In default mode, no options are exposed — the default settings are used automatically. These defaults have been tested to work well in most scenarios.

Advanced Options

Scenario

Default: telephonic. Diarization scenario to optimize for. Options: telephonic, meeting, general.

Note: The default telephonic scenario has been tested to produce the best results in most cases and is not recommended to change.

Device

Default: auto. Processing device to use.

  • auto — Automatically selects the best available device
  • mps — Apple Silicon GPU (macOS only)
  • cuda — NVIDIA GPU (Linux/Windows only)
  • cpu — CPU only

Speaker Assignment Segment Anchor

Default: start. Segment anchor point for speaker assignment. Options: start, center, or end.

Speaker Majority Threshold

Default: 0.75. Minimum percentage (0.0–1.0) for majority voting speaker correction at sentence level.

Speaker Assignment Tolerance

Default: 1000,5000. Time tolerance for speaker assignment in milliseconds. Can be a single value (e.g., 2000) or a pair (e.g., 1000,5000).

Smooth speaker turns

Enabled by default. Enable smoothing to merge short speaker islands after majority voting. This reduces fragmentation and improves readability by reassigning short "island" turns (segments with few words surrounded by another speaker) to neighboring speakers.

  • Smooth Word Threshold — Default: 10. Maximum words in an island slice eligible for smoothing.
  • Smooth Ratio — Default: 2.0. Dominance ratio required for the neighbor turn to absorb the island.

Preprocess Audio

Disabled by default. When enabled, cuts excluded time ranges from the audio before processing. When disabled (default), the full audio is processed and excluded ranges are filtered from the RTTM output instead.

Exclude Time

Default: empty. Time ranges to exclude from diarization (e.g., 1:30-2:45,5:00-6:30). You can enter ranges manually or use the shared exclusion block. See Timeline Exclusions for details.

Extra Args

Default: empty. Additional arguments passed to nemo_diarize for advanced users who need to fine-tune behavior not exposed in the UI.

Related