view yeti/timefreq.yeti @ 135:8db5e4ab56ce

Ground-truth data in CSV and lab format, converted from the MIDI using Sonic Visualiser and then to lab using the script here

module timefreq;

// Obtain the time-frequency representation (based on constant-Q
// transform) as transcription input

af = load may.stream.audiofile;
mat = load may.matrix;
plot = load may.plot;
vec = load may.vector;

{ pow } = load may.mathmisc;

{ resampledTo } = load may.stream.resample;

{ cqt } = load cqt;

prepareTimeFrequency wavfile =
   (stream = resampledTo 44100 (af.openMono wavfile);

    streamLength =
        case stream.available of
        Known n: n;
        _: failWith "Audio file length unknown?!";
        esac;
      
    eprintln "streamLength = \(streamLength)";

   //!!! original also scales to peak = 0.5

    cq = cqt {
        maxFreq = stream.sampleRate / 3,
        minFreq = 27.5,
        binsPerOctave = 60
    } stream;

    //!!! note: original also modifies the Q and atomHopFactor
    eprintln "atomSpacing = \(cq.kernel.atomSpacing)";

    matrices = case cq.output (Spectrogram ()) of
        Real mm: mm;
        _: failWith "Expected real";
        esac;

    eprintln "have \(length matrices) matrices of size \(mat.size (head matrices)), isRowMajor? = \(mat.isRowMajor? (head matrices))";

    levels = concatMap do m:
        map do c: vec.sum c done (mat.asColumns m);
    done matrices;

    nztail = find (> 0.1) levels;
    nzonly = reverse (find (> 0.1) (reverse nztail));

    eprintln "non-zero columns start at \(length levels - length nztail), go on for \(length nzonly) [of \(length levels)]";
    
    nzstart = (length levels - length nztail) * cq.kernel.atomSpacing;
    nzduration = (length nzonly) * cq.kernel.atomSpacing;

    // Get a stream of columns at 25 per second.
    // 
    // The original picks samples at a rate of 100-per-second then
    // median filters to reduce noise then picks samples again at
    // 25-per-second. We don't do that (yet)

    samplesPerCol = stream.sampleRate / 25;
    var sample = samplesPerCol - nzstart;

    columns = take (nzduration / samplesPerCol)
       (concatMap do m:
            concatMap do col:
                sample := sample + cq.kernel.atomSpacing;
                if sample >= samplesPerCol then
                    sample := sample - samplesPerCol;
                    [col]
                else 
                    []
                fi;
            done (mat.asColumns m);
        done matrices);

    eprintln "have \(length columns) columns of \(vec.length (head columns)) values each";

    // drop the lowest 55 of the 600 bins
    columns = map do c:
        vec.slice c 55 (vec.length c);
    done columns;

    eprintln "now have \(length columns) columns of \(vec.length (head columns))";

//    plot.plot [ Grid (mat.fromColumns columns) ];

    columns);


{
    prepareTimeFrequency
}