import { DEFAULT_SESSION_OPTIONS, type OrtRuntime } from "./ort-runtime.js";
export interface EmbeddingRequest {
    /** Caller-defined id. Echoed back in the result map. */
    id: number;
    /** Row-major fbank (numFrames, 80). */
    fbank: Float64Array | Float32Array;
    numFrames: number;
}
/**
 * WeSpeaker ResNet34 backbone wrapper.
 *
 * Input: Kaldi fbank features (numFrames, 80) — already mean-normalized.
 * Output: 256-d speaker embedding.
 */
export declare class EmbeddingModel {
    private readonly session;
    private readonly ort;
    private constructor();
    static load(modelSource: string | ArrayBufferLike | Uint8Array, ort: OrtRuntime, sessionOptions?: typeof DEFAULT_SESSION_OPTIONS): Promise<EmbeddingModel>;
    /** Single-input inference. Convenience wrapper around `embedBatch`. */
    embed(fbank: Float64Array | Float32Array, numFrames: number, numBins?: number): Promise<Float32Array>;
    /**
     * Batched inference over heterogeneous-length inputs.
     *
     * Strategy: sort by descending length, slice into batches of `batchSize`,
     * pad each batch's inputs (zero-pad on the right) to the batch's max length.
     * Padding biases the WeSpeaker statistics pooling slightly, but bucketing
     * by length keeps the bias localized; the cost is small (< 0.1% DER drift
     * observed in practice).
     */
    embedBatch(requests: EmbeddingRequest[], options?: {
        batchSize?: number;
        numBins?: number;
        onProgress?: (processed: number, total: number) => void;
        /** Sleep N ms between consecutive batches. Throttle for cooler CPU/GPU. */
        batchYieldMs?: number;
    }): Promise<Map<number, Float32Array>>;
}
export declare function cosineSimilarity(a: Float32Array | number[], b: Float32Array | number[]): number;
//# sourceMappingURL=embedding.d.ts.map