view yeti/cqt.yeti @ 11:d7af615c2cc1

Comments

module cqt;

cqtkernel = load cqtkernel;
resample = load may.stream.resample;
manipulate = load may.stream.manipulate;
syn = load may.stream.syntheticstream;
cm = load may.matrix.complex;
mat = load may.matrix;
framer = load may.stream.framer;
cplx = load may.complex;
fft = load may.transform.fft;
vec = load may.vector;

{ pow, round, floor, ceil, log2, nextPowerOfTwo } = load may.mathmisc;

cqt str =
   (sampleRate = str.sampleRate;
    maxFreq = sampleRate/2;
    minFreq = 40;
    binsPerOctave = 24;

println "Here";

    octaves = ceil (log2 (maxFreq / minFreq));

println "Here: about to calculate stuff with \(octaves)";

    actualMinFreq = (maxFreq / (pow 2 octaves)) * (pow 2 (1/binsPerOctave));

    println "sampleRate = \(sampleRate), maxFreq = \(maxFreq), minFreq = \(minFreq), actualMinFreq = \(actualMinFreq), octaves = \(octaves), binsPerOctave = \(binsPerOctave)";

    kdata = cqtkernel.makeKernel { sampleRate, maxFreq, binsPerOctave };

    println "atomsPerFrame = \(kdata.atomsPerFrame)";

    streams = manipulate.duplicated octaves str;

    //!!! can't be right!
    kernel = cm.transposed (cm.conjugateTransposed kdata.kernel);

    println "have kernel";

    fftFunc = fft.forward kdata.fftSize;

    cqblocks =
        map do octave:
            frames = framer.monoFrames //!!! mono for now
                { framesize = kdata.fftSize, hop = kdata.fftHop }
                (resample.decimated (pow 2 octave) streams[octave]);
            map do frame:
                freq = fftFunc (cplx.complexArray frame (vec.zeros kdata.fftSize));
                cm.product kernel (cm.newComplexColumnVector freq);
            done frames;
        done [0..octaves-1];

    // cqblocks is now a list<list<matrix>>. Each top-level list
    // corresponds to an octave, from highest to lowest, each having
    // twice as many elements in its list as the next octave. The
    // sub-lists are sampled in time with an effective spacing of
    // fftSize * 2^(octave-1) audio frames, and the matrices are row
    // vectors with atomsPerFrame * binsPerOctave complex elements.


    //!!! The comment below isn't true -- this is based on traditional
    // cqt -- the atom hop structure for this version is different
    // though so review later on

    // Each (2^(octaves-1) * fftHop) input frames gives us an output
    // structure conceptually like this:
    //
    // [][][][][][][][]   <- fftHop frames per highest-octave output value
    // [][][][][][][][]      layered as many times as binsPerOctave (here 2)
    // [--][--][--][--]   <- fftHop*2 frames for the next lower octave
    // [--][--][--][--]      etc
    // [------][------]
    // [------][------]
    // [--------------]
    // [--------------]
    //
    // Our input is as a set of lazy frame lists, one list per octave.
    // For each frame in the lowest octave

    println "prepared block list";
    cqblocks;

    );

testStream = manipulate.withDuration 96000 (syn.sinusoid 48000 500);

println "have test stream";

cqt testStream;