import type { ResolvedDevice } from "../device.js";
import type { PcmAudio } from "../platform.js";
import type { VoiceEmbedding, VoiceTensor } from "../voice/types.js";
/** Everything an engine needs to render one chunk of speech. */
export interface SynthesisRequest {
    /** Already normalized text — engines do not normalize again. */
    readonly text: string;
    /** Speaker to render the text with. */
    readonly voice: VoiceEmbedding;
    /** Playback rate multiplier; `1` is the engine's natural pace. */
    readonly speed: number;
    /**
     * How expressive the delivery should be, overriding whatever default the
     * engine was constructed with.
     *
     * Named for the effect rather than for any one model's parameter: engines
     * map it onto their own control (Chatterbox calls it `exaggeration`), and
     * engines with no such control ignore it.
     */
    readonly expressiveness?: number;
    /**
     * Abandon the render when this aborts.
     *
     * Engines that cannot interrupt work in progress may ignore it — the caller
     * stops waiting either way. Engines that can should stop promptly, because
     * a render nobody is waiting for still occupies the one slot the worker
     * serialises calls through.
     */
    readonly signal?: AbortSignal;
}
/**
 * What an engine extracts from reference audio.
 *
 * Engines whose speaker representation is a single vector may return that
 * `Float32Array` directly; models that need several tensors (Chatterbox keeps
 * four) return them alongside it.
 */
export interface EmbedResult {
    readonly vector: Float32Array;
    readonly tensors?: Readonly<Record<string, VoiceTensor>>;
}
/**
 * The seam between the VoxShot facade and an actual model.
 *
 * Swapping in an ONNX Runtime Web backend means implementing this interface —
 * nothing above it needs to change.
 */
export interface SynthesisEngine {
    /** Stable identifier, useful for logs and cache keys. */
    readonly name: string;
    /** Sample rate of the audio {@link SynthesisEngine.synthesize} returns. */
    readonly sampleRate: number;
    /** Prepare weights for the given device. Called once before use. */
    load(device: ResolvedDevice): Promise<void>;
    /** Extract a speaker embedding from mono reference audio. */
    embed(audio: PcmAudio): Promise<Float32Array | EmbedResult>;
    /** Render one chunk of speech. */
    synthesize(request: SynthesisRequest): Promise<Float32Array>;
    /** Release any resources held by the engine. */
    dispose(): Promise<void>;
}
//# sourceMappingURL=types.d.ts.map