Can we compute a mel-scaled spectrogram? about meyda

[hughrawlinson]

Absolutely! The first step would be to do some research and find a few different sources for the definition of this feature. We've found that often, different implementations of features do slightly different things or use slightly different terminology. We want to be approachable for developers coming from other tools, and try to keep the terminology we use similar to what other tools do. I'd suggest having a look at YAAFE, librosa, and essentia. Additionally, if there are any academic papers on the subject, those would be useful too. Let me know what you find and we can move on to the next step 😄

[hughrawlinson]

Let's keep this issue open to track the implementation?

[hughrawlinson]

I don't think we support that right now, but maybe we could - we support MFCCs, so I would think those are related. Are you interested in working on that?

[vemusharan]

Thanks a lot @hughrawlinson for quick response. Yes, I would like to work on that. Could you please guide me :)

[chenp6]

I'm not sure whether my thought is right or not.
But I guess that the extractMelBands in node_modules\meyda\dist\node\main.js (for CLI and Nodejs) and node_modules\meyda\dist\web\meyda.js (for web) is where compute a mel scaled spectrogram.
However it is log scaled. So I just copy the extractMelBands , comment out the line melBands[i] = Math.log(melBands[i] + 1);for log-transform, and change rename it into extractUnloggedMelBands

function extractUnloggedMelBands(_a) {
       var ampSpectrum = _a.ampSpectrum, melFilterBank = _a.melFilterBank, bufferSize = _a.bufferSize;
       if (!ampSpectrum || !melFilterBank) {
           throw new TypeError("Valid ampSpectrum and melFilterBank are required to generate melBands");
       }
       var powSpec = extractPowerSpectrum({ ampSpectrum: ampSpectrum });
       var numFilters = melFilterBank.length;
       var filtered = new Array(numFilters).fill().map(() => new Float32Array(bufferSize / 2));
       var melBands = new Float32Array(numFilters);
   
       for (var i = 0; i < numFilters; i++) {
           for (var j = 0; j < bufferSize / 2; j++) {
               filtered[i][j] = melFilterBank[i][j] * powSpec[j];
               melBands[i] += filtered[i][j];
           }
           // Skip the log transformation
           // melBands[i] = Math.log(melBands[i] + 1); // commented out for unlogged version
       }
       return Array.from(melBands);
}

And then add a new config for it

 var extractors = /*#__PURE__*/Object.freeze({
        __proto__: null,
        amplitudeSpectrum: amplitudeSpectrum,
        buffer: buffer,
        chroma: chroma,
        complexSpectrum: complexSpectrum,
        energy: energy,
        loudness: loudness,
        melBands: extractMelBands,
        unloggedMelBands: extractUnloggedMelBands, // here!
       .....

Finally, you can use features.unloggedMelBandsto call it

const analyzer = Meyda.createMeydaAnalyzer({
    "audioContext": audioContext,
    "source": source,
    "bufferSize": 512,
    "featureExtractors": ["rms", "mfcc","unloggedMelBands"],
    "numberOfMFCCCoefficients":128,
    "callback": features => {
            console.log(features.unloggedMelBands);
    }
});