/**
 * VBx clustering. Direct port of pyannote.audio/utils/vbx.py (4.0.4),
 * which itself is derived from BUT VBx (Landini et al., 2022) without HMM.
 *
 * Notation matches the original Python; comments reference equation numbers
 * from the paper "Bayesian HMM clustering of x-vector sequences (VBx)".
 */
import type { Mat } from "../math/linalg.js";
export interface VBxResult {
    /** (T, S) responsibilities (marginal posteriors per frame and speaker). */
    gamma: Mat;
    /** (S,) ML-learned speaker priors. Speakers with pi ~ 0 were pruned by the model. */
    pi: Float64Array;
    /** ELBO trajectory (one entry per VB iteration). */
    elbo: number[];
}
export interface VBxOptions {
    Fa?: number;
    Fb?: number;
    maxIters?: number;
    epsilon?: number;
    /** AHC initialization labels, length T (values in 0..K-1). */
    ahcInit: Int32Array;
    /** Smoothing applied to one-hot AHC init via softmax(qinit * smoothing). Use < 0 to disable. */
    initSmoothing?: number;
}
/**
 * VBx core algorithm. PLDA assumes zero mean, diagonal across-class covariance Phi,
 * identity within-class covariance.
 *
 * X     - (T, D) PLDA-transformed feature vectors (one per chunk-speaker)
 * Phi   - (D,)   between-class diagonal covariance (= plda.psi[:lda_dim])
 */
export declare function vbx(X: Mat, Phi: Float64Array, opts: VBxOptions): VBxResult;
/**
 * pyannote-style cluster_vbx wrapper.
 * Returns gamma (T, S) and pi (S,). Speakers with pi <= 1e-7 should be discarded.
 */
export declare function clusterVbx(fea: Mat, phi: Float64Array, ahcInit: Int32Array, options: {
    Fa: number;
    Fb: number;
    maxIters?: number;
    initSmoothing?: number;
}): VBxResult;
//# sourceMappingURL=vbx.d.ts.map