openvino-genai-node/types/pipelines/whisperPipeline.d.ts

OpenVINO™ GenAI pipelines for using from Node.js environment

import { WhisperPipeline as WhisperPipelineWrapper } from "../addon.js";
import { WhisperDecodedResults } from "../decodedResults.js";
import { WhisperGenerationConfig, WhisperPipelineProperties, StreamingStatus } from "../utils.js";
import { Tokenizer } from "../tokenizer.js";
export type RawSpeechInput = Float32Array | number[];
/**
 * Options for Whisper generation methods.
 */
export type WhisperGenerateOptions = {
    /** Generation configuration (e.g. language, task, return_timestamps). */
    generationConfig?: WhisperGenerationConfig;
    /** Callback invoked with each decoded text chunk; return StreamingStatus to control generation. */
    streamer?: (chunk: string) => StreamingStatus;
};
/**
 * Pipeline for automatic speech recognition using Whisper models.
 *
 * Expects raw audio normalized to approximately [-1, 1] at 16 kHz sample rate.
 * Use a WAV file or decode audio to Float32Array before calling generate().
 */
export declare class WhisperPipeline {
    protected readonly modelPath: string;
    protected readonly device: string;
    protected pipeline: WhisperPipelineWrapper | null;
    protected readonly properties: WhisperPipelineProperties;
    /**
     * Construct a Whisper pipeline from a folder containing model IRs and tokenizer.
     * @param modelPath - Path to the folder with model IRs and tokenizer (e.g. openvino_encoder_model.xml, preprocessor_config.json).
     * @param device - Inference device (e.g. "CPU", "GPU").
     * @param properties - Device and pipeline properties (e.g. word_timestamps: true, CACHE_DIR: "cache").
     */
    constructor(modelPath: string, device: string, properties?: WhisperPipelineProperties);
    /**
     * Load the pipeline. Must be called once before generate().
     */
    init(): Promise<void>;
    /**
     * Stream speech recognition results as an async iterator.
     * The iterator yields decoded text chunks during generation.
     * When generation finishes, the full decoded text is returned as the final
     * iterator value (`done: true`). This value is not available through
     * `for await...of`; call `next()` directly to read it.
     *
     * For custom streaming control, use {@link generate} with a streamer callback instead.
     *
     * @param rawSpeech - Audio samples as Float32Array or number[], normalized to ~[-1, 1], 16 kHz.
     * @param options - Optional generation config (e.g. language, task, return_timestamps).
     * @returns Async iterator that yields decoded text chunks as strings.
     */
    stream(rawSpeech: RawSpeechInput, options?: WhisperGenerateOptions): AsyncIterableIterator<string>;
    /**
     * Run speech recognition with optional streaming.
     *
     * For simple streaming use cases, consider using {@link stream}, which provides
     * a convenient async iterator interface.
     *
     * @param rawSpeech - Audio samples as Float32Array or number[], normalized to ~[-1, 1], 16 kHz.
     * @param options - Optional parameters.
     * @param options.generationConfig - Generation config (e.g., language, task, return_timestamps).
     * @param options.streamer - Optional callback invoked for each decoded chunk.
     * - Return a `StreamingStatus` flag to indicate whether generation should be stopped or cancelled
     * @returns Decoded texts, scores, optional chunks with timestamps, and perf metrics.
     */
    generate(rawSpeech: RawSpeechInput, options?: WhisperGenerateOptions): Promise<WhisperDecodedResults>;
    /**
     * Get the pipeline tokenizer.
     */
    getTokenizer(): Tokenizer;
    /**
     * Get current generation config (language, task, return_timestamps, etc.).
     */
    getGenerationConfig(): Partial<WhisperGenerationConfig>;
    /**
     * Update generation config (e.g. language, task, return_timestamps).
     */
    setGenerationConfig(config: WhisperGenerationConfig): void;
}