speech-detector/dist/silero.js

A node.js library for detecting speech in audio using Silero's VAD model

import ort from "onnxruntime-web";
import * as fs from "fs";
// This resolves a bug with WASM in nodejs.
ort.env.wasm.numThreads = 1;
ort.env.remoteModels = false;
class Silero {
    session;
    sampleRate;
    // Internal variables
    c;
    h;
    constructor(session, sampleRate) {
        this.session = session;
        // Note: I suppose the model requires a tensor as input.
        this.sampleRate = new ort.Tensor("int64", [sampleRate]);
        // Note: No idea what these inputs to the model are for...
        this.c = new ort.Tensor("float32", Array(2 * 64).fill(0), [2, 1, 64]);
        this.h = new ort.Tensor("float32", Array(2 * 64).fill(0), [2, 1, 64]);
    }
    static async create(sampleRate, uri = "./model/silero_vad.onnx") {
        const opt = {
            executionProviders: ["wasm"],
            logSeverityLevel: 3,
            logVerbosityLevel: 3,
        };
        // For compatability convert the URI into a properly
        // formatted URL. This will work for NodeJS and Web.
        const path = new URL(uri, import.meta.url);
        let session;
        if (typeof window === "undefined") {
            // Only read in the model file in NodeJS.
            const model = fs.readFileSync(path);
            session = await ort.InferenceSession.create(model, opt);
        }
        else {
            session = await ort.InferenceSession.create(uri, opt);
        }
        return new Silero(session, sampleRate);
    }
    async process(audio) {
        const t = new ort.Tensor("float32", audio, [1, audio.length]);
        const input = {
            input: t,
            h: this.h,
            c: this.c,
            sr: this.sampleRate,
        };
        const output = await this.session.run(input);
        this.h = output.hn;
        this.c = output.cn;
        return output.output.data[0];
    }
}
export { Silero };