Question 54118 in Javascript Question posted in
A very good W3school tutorial can be found here.

How can I read audio wav in javascript and transform to [-1, 1]

I want to use the yamnet TensorFlow model, but it needs to tensor between [-1, 1] to enter in the model.

How can I do that? I want to transform it to [-1, 1]?

Option #1

const readWavAudio = async () => {
  wavfiles = await fs.readFileSync('archive/dataset/1/1_484.wav')
  const wav = new wavefile.WaveFile();
  wav.fromBuffer(wavfiles)
  wav.toSampleRate(16000);
  return await wav.getSamples(false, Float32Array)
}

Option #2

async function readbuf_async() {
  data = fs.readFileSync("teste.wav");
  buffer = [];

  for (o = 0; o < data.length - 4 * 1; o += 4 * 1) {
    buffer.push(data.readFloatBE(o));
  }

  return buffer;
}
Tags:

2

Answers


  1. 0

    You can scale the vector to a range between [-1, 1] by utilizing linear interpolation:

    const normalizeVector = (vector) => {
  const { min, max } = vectorScale(vector),
    originalScale = [min, max],
    adjustedScale = [-1, 1];
  return vector.map((value) => scaleValue(value, originalScale, adjustedScale));
};

const scaleValue = (value, originalScale, adjustedScale) => {
  const [minO, maxO] = originalScale,
    [minA, maxA] = adjustedScale;
  return ((value - minO) * (maxA - minA)) / (maxO - minO) + minA;
};

const vectorScale = (vector) =>
  vector.reduce(
    (scale, value) => {
      if (value < scale.min) scale.min = value;
      if (value > scale.max) scale.max = value;
      return scale;
    },
    { min: Number.MAX_SAFE_INTEGER, max: Number.MIN_SAFE_INTEGER }
  );

    Full example

    Here is a Node.js module that creates a JSON data file containing the desired values.

    import fs from "fs";
import path from "path";
import wavefile from "wavefile";

const main = async () => {
  await analyzeAudioFile("archive/dataset/1/1_484.wav");
  console.log("Finished...");
};

const analyzeAudioFile = async (filename) => {
  console.log("Analyzing audio...");
  const samples = await readWavAudio(filename),
    scaled = normalizeVector(Array.from(samples)),
    outputFilename = `${fileNameFromPath(filename)}.json`;
  fs.writeFileSync(outputFilename, JSON.stringify(scaled));
  console.log(`Wrote ${outputFilename}`);
};

const readWavAudio = async (wavFilename) => {
  const wavfiles = fs.readFileSync(wavFilename);
  const wav = new wavefile.WaveFile();
  wav.fromBuffer(wavfiles);
  wav.toSampleRate(16000);
  return wav.getSamples(false, Float32Array);
};

const normalizeVector = (vector) => {
  const { min, max } = vectorScale(vector),
    originalScale = [min, max],
    adjustedScale = [-1, 1];
  return vector.map((value) => scaleValue(value, originalScale, adjustedScale));
};

const scaleValue = (value, originalScale, adjustedScale) => {
  const [minO, maxO] = originalScale,
    [minA, maxA] = adjustedScale;
  return ((value - minO) * (maxA - minA)) / (maxO - minO) + minA;
};

const vectorScale = (vector) =>
  vector.reduce(
    (scale, value) => {
      if (value < scale.min) scale.min = value;
      if (value > scale.max) scale.max = value;
      return scale;
    },
    { min: Number.MAX_SAFE_INTEGER, max: Number.MIN_SAFE_INTEGER }
  );

const fileNameFromPath = (filename) => {
  const extension = path.extname(filename);
  return path.basename(filename, extension);
};

main();

    Output

    The Node.js module script will create a JSON file that contains an array of values between [-1, 1].

    Analyzing audio...
Wrote 1_484.json
Finished...

    Optimization

    Here is a slightly optimized version of normalizeVector:

    Note: The 2 is the result of min - max of the adjusted scale i.e. 1 - (-1).

    const normalizeVector = (vector) => {
  const { min, max } = vectorScale(vector),
    range = max - min;
  return vector.map((value) => lerp(value, range, 2, min, -1));
};

const lerp = (value, range, newRange, min, newMin) =>
  ((value - min) * newRange) / range + newMin;
    Login or Signup to reply.
  2. 0

    I was able to make use of wavefile for this. In my case, I was starting with base64 encoded mulaw audio from Twilio. Dividing by 32768 (the max 16-bit signed integer) is what gets you values between [-1, 1].

    const WaveFile = require('wavefile').WaveFile;
const tf = require('@tensorflow/tfjs');

const wav = new WaveFile();
wav.fromScratch(1, 8000, '8m', Buffer.from(msg.media.payload, 'base64'));
wav.fromMuLaw();
wav.toSampleRate(16000);
let samples = wav.getSamples(false, Int16Array);
samples = Array.from(samples).map(s => s / 32768);
waveform = tf.tensor(samples)
    Login or Signup to reply.
Please signup or login to give your own answer.

Back To Top
Search