Chris@12: 
Chris@12: module timefreq;
Chris@12: 
Chris@12: // Obtain the time-frequency representation (based on constant-Q
Chris@12: // transform) as transcription input
Chris@12: 
Chris@12: af = load may.stream.audiofile;
Chris@12: mat = load may.matrix;
Chris@12: plot = load may.plot;
Chris@12: vec = load may.vector;
Chris@12: 
Chris@12: { pow } = load may.mathmisc;
Chris@12: 
Chris@12: { resampledTo } = load may.stream.resample;
Chris@12: 
Chris@12: { cqt } = load cqt;
Chris@12: 
Chris@12: prepareTimeFrequency wavfile =
Chris@12:    (stream = resampledTo 44100 (af.openMono wavfile);
Chris@12: 
Chris@12:     streamLength =
Chris@12:         case stream.available of
Chris@12:         Known n: n;
Chris@12:         _: failWith "Audio file length unknown?!";
Chris@12:         esac;
Chris@12:       
Chris@12:     eprintln "streamLength = \(streamLength)";
Chris@12: 
Chris@12:    //!!! original also scales to peak = 0.5
Chris@12: 
Chris@12:     cq = cqt {
Chris@12:         maxFreq = stream.sampleRate / 3,
Chris@12:         minFreq = 27.5,
Chris@12:         binsPerOctave = 60
Chris@12:     } stream;
Chris@12: 
Chris@12:     //!!! note: original also modifies the Q and atomHopFactor
Chris@12:     eprintln "atomSpacing = \(cq.kernel.atomSpacing)";
Chris@12: 
Chris@12:     matrices = case cq.output (Spectrogram ()) of
Chris@12:         Real mm: mm;
Chris@12:         _: failWith "Expected real";
Chris@12:         esac;
Chris@12: 
Chris@12:     eprintln "have \(length matrices) matrices of size \(mat.size (head matrices)), isRowMajor? = \(mat.isRowMajor? (head matrices))";
Chris@12: 
Chris@12:     levels = concatMap do m:
Chris@12:         map do c: vec.sum c done (mat.asColumns m);
Chris@12:     done matrices;
Chris@12: 
Chris@12:     nztail = find (> 0.1) levels;
Chris@12:     nzonly = reverse (find (> 0.1) (reverse nztail));
Chris@12: 
Chris@12:     eprintln "non-zero columns start at \(length levels - length nztail), go on for \(length nzonly) [of \(length levels)]";
Chris@12:     
Chris@12:     nzstart = (length levels - length nztail) * cq.kernel.atomSpacing;
Chris@12:     nzduration = (length nzonly) * cq.kernel.atomSpacing;
Chris@12: 
Chris@12:     // Get a stream of columns at 25 per second.
Chris@12:     // 
Chris@12:     // The original picks samples at a rate of 100-per-second then
Chris@12:     // median filters to reduce noise then picks samples again at
Chris@12:     // 25-per-second. We don't do that (yet)
Chris@12: 
Chris@12:     samplesPerCol = stream.sampleRate / 25;
Chris@12:     var sample = samplesPerCol - nzstart;
Chris@12: 
Chris@12:     columns = take (nzduration / samplesPerCol)
Chris@12:        (concatMap do m:
Chris@12:             concatMap do col:
Chris@12:                 sample := sample + cq.kernel.atomSpacing;
Chris@12:                 if sample >= samplesPerCol then
Chris@12:                     sample := sample - samplesPerCol;
Chris@12:                     [col]
Chris@12:                 else 
Chris@12:                     []
Chris@12:                 fi;
Chris@12:             done (mat.asColumns m);
Chris@12:         done matrices);
Chris@12: 
Chris@12:     eprintln "have \(length columns) columns of \(vec.length (head columns)) values each";
Chris@12: 
Chris@12:     // drop the lowest 55 of the 600 bins
Chris@12:     columns = map do c:
Chris@12:         vec.slice c 55 (vec.length c);
Chris@12:     done columns;
Chris@12: 
Chris@12:     eprintln "now have \(length columns) columns of \(vec.length (head columns))";
Chris@12: 
Chris@12: //    plot.plot [ Grid (mat.fromColumns columns) ];
Chris@12: 
Chris@12:     columns);
Chris@12: 
Chris@12: 
Chris@12: {
Chris@12:     prepareTimeFrequency
Chris@12: }
Chris@12: