export interface FbankConfig {
    numMelBins: number;
    frameLengthMs: number;
    frameShiftMs: number;
    sampleFrequency: number;
    /** "hamming" only for now (matches WeSpeaker / pyannote). */
    windowType: "hamming";
    preemphasisCoefficient: number;
    removeDcOffset: boolean;
    roundToPowerOfTwo: boolean;
    lowFreq: number;
    /** 0 means Nyquist (sampleFrequency / 2). */
    highFreq: number;
    usePower: boolean;
    useLogFbank: boolean;
    /** Float32 epsilon, used by torchaudio for the log floor. */
    epsilon: number;
}
export declare const PYANNOTE_FBANK_CONFIG: FbankConfig;
/**
 * Kaldi-compatible fbank features. Mirrors torchaudio.compliance.kaldi.fbank
 * with the parameter set used by the WeSpeaker / pyannote pipeline.
 *
 * Returns row-major (numFrames, numMelBins) as a single Float64Array, plus shape.
 */
export declare function kaldiFbank(waveform: Float32Array | Float64Array, config?: FbankConfig): {
    data: Float64Array;
    numFrames: number;
    numBins: number;
};
/**
 * Pyannote / WeSpeaker fbank: Kaldi fbank + per-bin mean normalization across time.
 * This is what the embedding ResNet34 expects as input.
 */
export declare function pyannoteFbank(waveform: Float32Array | Float64Array, config?: FbankConfig): {
    data: Float64Array;
    numFrames: number;
    numBins: number;
};
//# sourceMappingURL=fbank.d.ts.map