Speaker labels and voice matching
Speaker labels help readers follow a transcript and check who accepted an action. Review important attributions against the recording: overlapping speech, microphones and background noise can affect the result.
Labels supplied by a meeting platform or transcription provider are distinct from matching a person's voice against a saved biometric template. A displayed participant name alone does not establish whose voice is in an audio sample.
What MFCC features represent
MFCC means Mel-Frequency Cepstral Coefficients. These are numerical features calculated from an audio signal; the librosa reference documents the calculation. They can form part of an audio comparison, but a numerical template is not proof of a person's identity or guaranteed uniqueness.
A voice template and its source recording are different data objects. Keeping a template does not, by itself, establish that the source audio has been deleted or that the template is anonymous.
Requirements before biometric matching
A workflow needs a reliable link between the person, the applicable consent and that person's audio source. Cutting a time window from a mixed meeting recording does not isolate a single participant when voices overlap. A matching threshold cannot establish that link by itself.
Availability in ReVoice depends on enabled functionality and the applicable workflow. Do not assume that one voice sample enables recognition in every meeting or imported recording. Confirm the supported source, consent scope and current availability before relying on the feature.
Privacy and control
Biometric data used to uniquely identify a person falls within GDPR's special categories. The controller needs a legal basis and an applicable Article 9 condition. See IMY's guidance on sensitive personal data.
For ReVoice voice enrollment, review the explicit-consent text, stated purpose and scope, retention period and withdrawal route shown in the actual flow. Voice-profile controls are under Settings > Privacy & Data. Check the current privacy policy, relevant sub-processors and evidence of storage, access and deletion, including the source audio and backups.
Voice matching does not establish a person's emotion, personality or suitability for a job. Evaluate those questions independently of speaker identification.