{
 "metadata": {
  "name": "Musichackathon - Machinedness"
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [
    {
     "cell_type": "code",
     "collapsed": false,
     "input": "import os\nimport glob\nimport librosa\n\nfrom pprint import pprint\n\nimport numpy as np\nimport scipy\n\nimport functools\nspecshow = functools.partial(imshow, aspect='auto', interpolation='nearest', origin='lower')",
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 28
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": "SR    = 22050\nN_FFT = 1024\nN_MELS = 256\nHOP   = N_FFT / 8\nFMAX  = 8000",
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 3
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": "def load_features(f):\n    \n    y, sr = librosa.load(f, sr=SR)\n\n    S = librosa.feature.melspectrogram(y, sr=SR, n_fft=N_FFT, hop_length=HOP, fmax=FMAX, n_mels=N_MELS)\n    M = librosa.feature.mfcc(librosa.logamplitude(S))\n    \n    return S, M",
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 4
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": "def beat_stability(S):\n    \n    onset_strength = np.median(np.maximum(0.0, np.diff(S, axis=1)), axis=0)\n    onset_strength = onset_strength / onset_strength.max()\n    \n    tempo, beats = librosa.beat.beat_track(onsets=onset_strength, hop_length=HOP, sr=SR, trim=False)\n    \n    beat_times = librosa.core.frames_to_time(beats, sr=SR, hop_length=HOP)\n    \n    intervals = np.diff(beat_times)\n    \n    stability = 1.0 - np.std(intervals) / (tempo / 60)\n    \n    return stability, beats, tempo",
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 5
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": "def struct_rep(X, k=0.05, display=True):\n    \n    true_k = np.ceil(k * X.shape[1])\n    R = librosa.segment.recurrence_matrix(X, k=k, width=8)\n    S = librosa.segment.structure_feature(R).astype(np.float32)\n    S = scipy.signal.medfilt2d(S, np.array([1, 5]))\n    \n    num_repeats = np.sum(S, axis=0)\n    \n    if display:\n        figure(figsize=(8,8))\n        specshow(R, cmap='gray_r')\n        xlabel('Beat #')\n        ylabel('Beat #')\n        title('Beat-beat similarity')\n        \n        figure(figsize=(8,8))\n        subplot(211)\n        specshow(S, cmap='gray_r')\n        ylabel('Lag')\n        title('Beat-lag repetition')\n\n        subplot(212)\n        plot(num_repeats)\n        axis('tight')\n        xlabel('Beat #')\n        ylabel('# Repetitions')\n    \n    score = np.mean(num_repeats / true_k)\n    return score",
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 31
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": "def score_file(f):\n    S, M = load_features(f)\n    \n    stability_score, B, tempo = beat_stability(S)\n    M_beat = librosa.feature.sync(M, B, aggregate=np.mean)[1:,:]\n    repetition_score = struct_rep(M_beat)\n    return stability_score, repetition_score",
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 7
    }
   ],
   "metadata": {}
  }
 ]
}