import * as ort from "onnxruntime-web";
import type { ModelConfig } from "./model-types.js";
import type { AsrTranscriber, AudioSamples, ConfigureOrtWebOptions, DecoderOptions, OrtTensor, SessionOptions, TensorData, TensorMap, TokenFrame, TranscriptResult } from "./types.js";
import { EncoderModel } from "./models.js";
type TokenVocabulary = string[];
type EncoderLayout = "BCT" | "BTV" | "BVT";
type TensorMetadata = {
    name: string;
    type?: string;
    shape?: readonly (number | string | undefined)[];
};
type DecoderKind = "whisper-ort" | "whisper-hf" | "aed" | "gigaam-rnnt" | "tone-ctc" | "sherpa-transducer" | "ctc" | "rnnt" | "tdt" | "nemo-conformer";
interface DecoderResult {
    tokenIds: number[];
    tokenFrames: TokenFrame[];
    totalFrames: number;
}
interface EncodedAudio {
    encodedData: TensorData;
    encodedDims: readonly number[];
    encodedLayout: EncoderLayout;
    encodedLength: number;
}
interface EncoderInputs {
    signal: OrtTensor;
    length: OrtTensor;
}
interface EncoderLike {
    audioSignalName: string;
    lengthName: string;
    prepareInputsFromWaveform(samples: Float32Array): EncoderInputs;
    run(signal: OrtTensor, length: OrtTensor): Promise<EncodedAudio>;
}
interface PreprocessorLike {
    run(audioSamples: Float32Array): Promise<EncoderInputs>;
}
interface DecoderLike {
    decode(encodedData: TensorData, encodedDims: readonly number[], encodedLayout: EncoderLayout, encodedLength: number): Promise<DecoderResult>;
}
interface CanaryOptions {
    language?: string;
    targetLanguage?: string;
    pnc?: "yes" | "no";
}
interface AsrModelOptions {
    preprocessor?: PreprocessorLike | null;
    encoder: EncoderLike;
    decoder: DecoderLike;
    tokens: TokenVocabulary;
    sampleRate?: number;
}
interface SessionBundleOptions {
    config?: ModelConfig;
    tokens: TokenVocabulary;
    encoderSession: ort.InferenceSession;
    decoderSession: ort.InferenceSession;
}
interface GigaamModelOptions extends SessionBundleOptions {
    jointSession: ort.InferenceSession;
}
interface ToneModelOptions {
    config?: ModelConfig;
    tokens: TokenVocabulary;
    session: ort.InferenceSession;
}
interface SherpaModelOptions {
    config?: ModelConfig;
    tokens: TokenVocabulary;
    encoderSession: ort.InferenceSession;
    decoderSession: ort.InferenceSession;
    joinerSession: ort.InferenceSession;
}
interface CreateAsrModelOptions {
    modelType?: string;
    decoderKind?: DecoderKind | string;
    config?: ModelConfig;
    preprocessorModel?: string | null;
    encoderModel?: string | null;
    decoderModel?: string | null;
    decoderJointModel?: string | null;
    whisperModel?: string | null;
    vocabularyText?: string | null;
    vocabJson?: string | null;
    addedTokensJson?: string | null;
    sessionOptions?: SessionOptions;
    decoderOptions?: DecoderOptions;
}
/** Configure ONNX Runtime Web global environment settings. */
export declare function configureOrtWeb(options?: ConfigureOrtWebOptions): void;
export declare class AsrModel implements AsrTranscriber {
    readonly preprocessor: PreprocessorLike | null;
    readonly encoder: EncoderLike;
    readonly decoder: DecoderLike;
    readonly tokens: TokenVocabulary;
    readonly sampleRate: number;
    constructor({ preprocessor, encoder, decoder, tokens, sampleRate }: AsrModelOptions);
    transcribeSamples(samples: AudioSamples, sampleRate?: number): Promise<TranscriptResult>;
    transcribeWavBuffer(arrayBuffer: ArrayBuffer): Promise<TranscriptResult>;
}
export declare class NemoAedModel implements AsrTranscriber {
    readonly config?: ModelConfig;
    readonly tokens: TokenVocabulary;
    readonly tokenToId: Map<string, number>;
    readonly encoderSession: ort.InferenceSession;
    readonly decoderSession: ort.InferenceSession;
    readonly encoderHelper: EncoderModel;
    readonly sampleRate: number;
    readonly maxSequenceLength: number;
    readonly encoderOutputEmbeddingsName: string;
    readonly encoderOutputMaskName: string;
    readonly decoderInputIdName: string;
    readonly decoderEncoderEmbeddingsName: string;
    readonly decoderEncoderMaskName: string;
    readonly decoderMemsName: string;
    readonly decoderInputIdType: string;
    readonly decoderMemsType: string;
    readonly decoderMemsShapeTemplate: readonly (number | string | undefined)[];
    readonly logitsName: string;
    readonly decoderHiddenStatesName: string;
    constructor({ config, tokens, encoderSession, decoderSession }: SessionBundleOptions);
    canaryPrefix(options?: CanaryOptions): number[];
    initialMems(batchSize: number): ort.Tensor;
    transcribeSamples(samples: AudioSamples, sampleRate?: number, options?: CanaryOptions): Promise<TranscriptResult>;
    transcribeWavBuffer(arrayBuffer: ArrayBuffer, options?: CanaryOptions): Promise<TranscriptResult>;
}
export declare class GigaamRnntModel implements AsrTranscriber {
    readonly config?: ModelConfig;
    readonly tokens: TokenVocabulary;
    readonly encoderSession: ort.InferenceSession;
    readonly decoderSession: ort.InferenceSession;
    readonly jointSession: ort.InferenceSession;
    readonly sampleRate: number;
    readonly encoderHelper: EncoderModel;
    readonly blankTokenId: number;
    readonly maxTokensPerStep: number;
    readonly decoderTargetName: string;
    readonly decoderTargetLengthName: string | null;
    readonly decoderTargetType: string;
    readonly decoderTargetLengthType: string;
    readonly decoderStateInputNames: string[];
    readonly decoderVectorOutputName: string;
    readonly decoderStateOutputNames: string[];
    readonly jointEncoderInputName: string;
    readonly jointDecoderInputName: string;
    readonly jointOutputName: string;
    readonly jointEncoderShape: readonly (number | string | undefined)[];
    readonly jointDecoderShape: readonly (number | string | undefined)[];
    constructor({ config, tokens, encoderSession, decoderSession, jointSession }: GigaamModelOptions);
    initialDecoderStates(batchSize?: number): Map<string, ort.Tensor>;
    adaptJointInput(tensor: ort.Tensor, targetShape: readonly (number | string | undefined)[]): ort.Tensor;
    runJointArgmax(encoderFrameTensor: ort.Tensor, decoderVectorTensor: ort.Tensor): Promise<number>;
    transcribeSamples(samples: AudioSamples, sampleRate?: number): Promise<TranscriptResult>;
    transcribeWavBuffer(arrayBuffer: ArrayBuffer): Promise<TranscriptResult>;
}
export declare class ToneCtcModel implements AsrTranscriber {
    readonly config?: ModelConfig;
    readonly tokens: TokenVocabulary;
    readonly session: ort.InferenceSession;
    readonly sampleRate: number;
    readonly blankTokenId: number;
    readonly signalName: string;
    readonly stateName: string;
    readonly logitsName: string;
    readonly nextStateName: string;
    readonly signalMeta: TensorMetadata | null;
    readonly stateMeta: TensorMetadata | null;
    readonly chunkSamples: number;
    readonly stateType: string;
    readonly stateShape: number[];
    constructor({ config, tokens, session }: ToneModelOptions);
    toIntPcm(samples: Float32Array): Int32Array;
    decodeCtcGreedy(tokens: number[], tokenFrames: TokenFrame[], sequence: readonly number[], frameStartIndex: number): void;
    transcribeSamples(samples: AudioSamples, sampleRate?: number): Promise<TranscriptResult>;
    transcribeWavBuffer(arrayBuffer: ArrayBuffer): Promise<TranscriptResult>;
}
export declare class SherpaTransducerModel implements AsrTranscriber {
    readonly config: ModelConfig;
    readonly tokens: TokenVocabulary;
    readonly encoderSession: ort.InferenceSession;
    readonly decoderSession: ort.InferenceSession;
    readonly joinerSession: ort.InferenceSession;
    readonly sampleRate: number;
    readonly blankTokenId: number;
    readonly maxSymbolsPerFrame: number;
    readonly encoderInputName: string;
    readonly encoderLengthName: string | null;
    readonly encoderInputMeta: TensorMetadata | null;
    readonly encoderLengthMeta: TensorMetadata | null;
    readonly encoderOutputName: string;
    readonly encoderLengthOutName: string | null;
    readonly decoderInputName: string;
    readonly decoderInputMeta: TensorMetadata | null;
    readonly decoderOutputName: string;
    readonly contextSize: number;
    readonly decoderInputType: string;
    readonly joinerEncName: string;
    readonly joinerDecName: string;
    readonly joinerOutputName: string;
    readonly joinerEncShape: readonly (number | string | undefined)[];
    readonly joinerDecShape: readonly (number | string | undefined)[];
    readonly encoderFeatureConfig: ModelConfig;
    readonly encoderHelper: EncoderModel;
    constructor({ config, tokens, encoderSession, decoderSession, joinerSession }: SherpaModelOptions);
    adaptTensorToShape(tensor: ort.Tensor, targetShape: readonly (number | string | undefined)[]): ort.Tensor;
    makeDecoderInput(context: readonly number[]): ort.Tensor;
    runEncoder(samples: Float32Array): Promise<TensorMap>;
    transcribeSamples(samples: AudioSamples, sampleRate?: number): Promise<TranscriptResult>;
    transcribeWavBuffer(arrayBuffer: ArrayBuffer): Promise<TranscriptResult>;
}
/** Create an ASR model runtime from resolved model files and config. */
export declare function createAsrModel({ modelType, decoderKind, config, preprocessorModel, encoderModel, decoderModel, decoderJointModel, whisperModel, vocabularyText, vocabJson, addedTokensJson, sessionOptions, decoderOptions, }?: CreateAsrModelOptions): Promise<AsrTranscriber>;
export {};
//# sourceMappingURL=asr-model.d.ts.map