Trait Clusterer

Source

pub trait Clusterer: Send + Sync {
    // Required methods
    fn cluster(
        &self,
        embeddings: &[Vec<f32>],
    ) -> Result<Vec<usize>, ClustererError>;
    fn max_clusters(&self) -> usize;
}

Expand description

Speaker clusterer — turns a batch of L2-normalized speaker embeddings into per-embedding cluster labels in the range 0..K where K is the inferred number of clusters.

In v1.0 (M3) the polyvoice crate introduces Clusterer as the canonical trait. The legacy free functions ahc::agglomerative_cluster_auto and spectral::spectral_cluster remain available — M6 will deprecate them.

Required Methods§

Source

fn cluster(&self, embeddings: &[Vec<f32>]) -> Result<Vec<usize>, ClustererError>

Cluster embeddings. Each inner vector must have the same length and be approximately L2-normalized.

Requires: embeddings.len() >= 1. Guarantees on Ok: result.len() == embeddings.len(), result[i] < unique(result).count() (compact 0..K numbering).

Source