Expand description
Speaker embeddings: a fixed-length vector that says what a voice sounds like, independent of what it says.
Two pieces, both small: fbank, Kaldi’s log mel filterbank in pure Rust
(the features WeSpeaker’s models are trained on), and SpeakerEncoder,
the network on ONNX Runtime. Two recordings of one voice land close
together, measured by cosine; that is what lets an application find
the blend of Kokoro voices closest to a recording.
An embedding is a similarity, never an identity. It says how alike two voices sound, which a recording, a synthesizer or an impression can all influence; it is not evidence of who is speaking and must not be used to authenticate anyone.
Weights are not bundled: the loqui facade downloads the pinned model.
WeSpeaker’s VoxCeleb models are licensed CC-BY-4.0, the licence of the
VoxCeleb dataset; see NOTICE.
Re-exports§
pub use encoder::MIN_FRAMES;pub use encoder::SpeakerEncoder;pub use encoder::cosine;pub use encoder::normalize;pub use fbank::Fbank;pub use fbank::Features;pub use fbank::MEL_BINS;pub use fbank::SAMPLE_RATE;
Modules§
- encoder
- A speaker-embedding network on ONNX Runtime.
- fbank
- Kaldi’s log mel filterbank, as WeSpeaker computes it for its speaker models.
Enums§
- Error
- Why an embedding could not be computed.