meyda/dist/node/extractors/mfcc.js

Real-time feature extraction for the web audio api

'use strict';

function extractPowerSpectrum (_a) {
    var ampSpectrum = _a.ampSpectrum;
    if (typeof ampSpectrum !== "object") {
        throw new TypeError();
    }
    var powerSpectrum = new Float32Array(ampSpectrum.length);
    for (var i = 0; i < powerSpectrum.length; i++) {
        powerSpectrum[i] = Math.pow(ampSpectrum[i], 2);
    }
    return powerSpectrum;
}

function extractMelBands (_a) {
    var ampSpectrum = _a.ampSpectrum, melFilterBank = _a.melFilterBank, bufferSize = _a.bufferSize;
    if (typeof ampSpectrum !== "object") {
        throw new TypeError("Valid ampSpectrum is required to generate melBands");
    }
    if (typeof melFilterBank !== "object") {
        throw new TypeError("Valid melFilterBank is required to generate melBands");
    }
    var powSpec = extractPowerSpectrum({ ampSpectrum: ampSpectrum });
    var numFilters = melFilterBank.length;
    var filtered = Array(numFilters);
    var loggedMelBands = new Float32Array(numFilters);
    for (var i = 0; i < loggedMelBands.length; i++) {
        filtered[i] = new Float32Array(bufferSize / 2);
        loggedMelBands[i] = 0;
        for (var j = 0; j < bufferSize / 2; j++) {
            //point-wise multiplication between power spectrum and filterbanks.
            filtered[i][j] = melFilterBank[i][j] * powSpec[j];
            //summing up all of the coefficients into one array
            loggedMelBands[i] += filtered[i][j];
        }
        //log each coefficient.
        loggedMelBands[i] = Math.log(loggedMelBands[i] + 1);
    }
    return Array.prototype.slice.call(loggedMelBands);
}

function getDefaultExportFromCjs (x) {
	return x && x.__esModule && Object.prototype.hasOwnProperty.call(x, 'default') ? x['default'] : x;
}

/*===========================================================================*\
 * Discrete Cosine Transform
 *
 * (c) Vail Systems. Joshua Jung and Ben Bryan. 2015
 *
 * This code is not designed to be highly optimized but as an educational
 * tool to understand the Mel-scale and its related coefficients used in
 * human speech analysis.
\*===========================================================================*/

var cosMap = null;

// Builds a cosine map for the given input size. This allows multiple input sizes to be memoized automagically
// if you want to run the DCT over and over.
var memoizeCosines = function(N) {
  cosMap = cosMap || {};
  cosMap[N] = new Array(N*N);

  var PI_N = Math.PI / N;

  for (var k = 0; k < N; k++) {
    for (var n = 0; n < N; n++) {
      cosMap[N][n + (k * N)] = Math.cos(PI_N * (n + 0.5) * k);
    }
  }
};

function dct$2(signal, scale) {
  var L = signal.length;
  scale = scale || 2;

  if (!cosMap || !cosMap[L]) memoizeCosines(L);

  var coefficients = signal.map(function () {return 0;});

  return coefficients.map(function (__, ix) {
    return scale * signal.reduce(function (prev, cur, ix_, arr) {
      return prev + (cur * cosMap[L][ix_ + (ix * L)]);
    }, 0);
  });
}
var dct_1 = dct$2;

var dct = dct_1;

var dct$1 = /*@__PURE__*/getDefaultExportFromCjs(dct);

function mfcc (_a) {
    // Tutorial from:
    // http://practicalcryptography.com/miscellaneous/machine-learning
    // /guide-mel-frequency-cepstral-coefficients-mfccs/
    // @ts-ignore
    var ampSpectrum = _a.ampSpectrum, melFilterBank = _a.melFilterBank, numberOfMFCCCoefficients = _a.numberOfMFCCCoefficients, bufferSize = _a.bufferSize;
    var _numberOfMFCCCoefficients = Math.min(40, Math.max(1, numberOfMFCCCoefficients || 13));
    var numFilters = melFilterBank.length;
    if (numFilters < _numberOfMFCCCoefficients) {
        throw new Error("Insufficient filter bank for requested number of coefficients");
    }
    var loggedMelBandsArray = extractMelBands({
        ampSpectrum: ampSpectrum,
        melFilterBank: melFilterBank,
        bufferSize: bufferSize,
    });
    var mfccs = dct$1(loggedMelBandsArray).slice(0, _numberOfMFCCCoefficients);
    return mfccs;
}

module.exports = mfcc;