/**
 * SpeakerEmbedding — speaker identity embeddings via the server's
 * /v1/audio/speaker_embeddings route.
 *
 * The browser SDK is server-route-only; there is no local C-ABI native
 * execution. This facade POSTs PCM-f32 audio bytes (as base64) to the
 * hosted native backend (sherpa-onnx ERes2NetV2) and returns a
 * normalized float-vector suitable for cosine similarity comparisons.
 *
 * Python reference: octomil-python/octomil/audio/speaker_embedding.py
 * Route: POST /v1/audio/speaker_embeddings (octomil-python/octomil/serve/app.py:1457)
 *
 * Request body:
 *   { audio_pcm_f32_base64: string, sample_rate_hz: number,
 *     model?: string, deadline_ms?: number }
 *
 * Response body:
 *   { object: "audio.speaker.embedding", model: string, sample_rate_hz: number,
 *     embedding: number[], dimensions: number }
 */
export interface SpeakerEmbeddingCreateRequest {
    /**
     * Raw mono PCM-f32 audio as ArrayBuffer / Uint8Array.
     * Bytes must be float32-LE.
     */
    audioBytes: ArrayBuffer | Uint8Array;
    /** Sample rate of the audio, in Hz (e.g. 16000). */
    sampleRateHz: number;
    /**
     * Model identifier. Defaults to "sherpa-eres2netv2-base" server-side.
     * Other model names will be rejected as UNSUPPORTED_MODALITY.
     */
    model?: string;
    /** Per-request deadline in milliseconds. Defaults to 300 000 server-side. */
    deadlineMs?: number;
}
export interface SpeakerEmbeddingCreateResponse {
    readonly object: "audio.speaker.embedding";
    /** Model that produced the embedding. */
    readonly model: string;
    readonly sampleRateHz: number;
    /**
     * L2-normalised float32 embedding vector.
     *
     * The canonical ERes2NetV2 base dimension is 512, but callers should NOT
     * hardcode 512 — read it from `dimensions` instead so future model updates
     * are handled transparently.
     */
    readonly embedding: number[];
    /** Embedding dimension (e.g. 512 for ERes2NetV2 base). */
    readonly dimensions: number;
}
export declare class SpeakerEmbedding {
    private readonly serverUrl;
    private readonly apiKey;
    constructor(serverUrl: string, apiKey: string);
    /**
     * Create a speaker identity embedding from a PCM-f32 audio clip.
     *
     * The returned vector is L2-normalised; compute cosine similarity with a
     * simple dot product:
     *   const score = embA.reduce((sum, v, i) => sum + v * (embB[i] ?? 0), 0);
     */
    create(request: SpeakerEmbeddingCreateRequest): Promise<SpeakerEmbeddingCreateResponse>;
}
//# sourceMappingURL=speaker-embedding.d.ts.map