Audio Analysis
Speaker Identification
Compares each clip's voiceprint with enrolled reference voices to name the speaker (or say Unknown), or to verify that it is the person a column claims.
When to use it
Tag known people in meetings, check a caller is the account holder, or find every recording of one presenter.
In a pipeline
Audio File InputDiarization (Who Spoke When)Speaker IdentificationExport CSV
Settings
| Setting | Type | Required | Default or hint |
|---|---|---|---|
| Mode | select | — | identify |
| Reference voices folder | text | Yes | files/voices — Anna.wav, or Anna/*.wav per person |
| Claimed speaker column | select | — | Select column… |
| Match when similarity is at least (0–1) | number | — | 0.86 |
| Model (Hugging Face id) | text | — | microsoft/wavlm-base-plus-sv |
| Run on | select | — | auto |
| Time limit (seconds) | number | — | 900 |
Example
Naming the speakers of each turn
- Mode=Identify: the closest enrolled speaker
- Reference voices folder=files/voices
| Speaker ID | Start (s) | Speaker | Similarity |
|---|---|---|---|
| Speaker 1 | 0.4 | Ana Ruiz | 0.93 |
| Speaker 2 | 13.1 | Unknown | 0.61 |
Worth knowing
- Enrol each person with Name.wav, or a folder per person (Ana Ruiz/*.wav) holding a few clips; 10–30 seconds of clean speech each is plenty.
- Similarity is the cosine of the voiceprints; 0.86 suits the default WavLM model. Raise it to be stricter.
- After Diarization, only each turn's stretch of the file is compared.
- Models run on the server in FlowDoo's Python and are downloaded once, on first use, into the data folder. A GPU is used when the server has one.