speaker-recognizer

To enroll a voice identity for speaker recognition, you provide a speaker name and a WAV audio sample to the enrollment HTTP API. The system uses a multi-turn, tag-based cluster system to build robust voice embeddings.

Yes, Telegram supports both voice enrollment and speaker recognition flows. The API-centric workflow links voice identities directly to Telegram inputs, handling unknown speaker turns and identity management.

Multi-turn voice enrollment builds embeddings by processing successive audio inputs through a cluster-tagging system. When an unknown speaker is detected, the enrollment decision matrix triggers additional audio capture turns to construct a robust voice identity.

The best way to manage voice identities is through the provided HTTP API. It supports listing enrolled voices, handling identity linking, and processing enrollment decision matrices for ongoing speaker recognition management.

You do not strictly need a microphone; the system supports both mic-based and Telegram-based enrollment. You can self-enroll by providing a name and a WAV sample to the API, which is processed through the reference flows.

A limitation of this self-enrollment system is that it requires structured, multi-turn audio inputs to build accurate embeddings. Single short audio clips may not satisfy the cluster-tagging requirements for a robust voice identity.