Run main.py with an input file name from data/test_chords/ with flag -m set to the method you want to use for detection, and -p for plotting the result. The default method is template matching. Example:\n\n```\npython3 main.py -i 'Grand Piano - Fazioli - major E middle.wav' -m hmm -p True\n```\n\nFor help, run `python3 main.py -h`\n
\n" }, { "path": "chromagram.py", "content": "\"\"\"\nAlgorithm based on the paper 'Automatic Chord Recognition from\nAudio Using Enhanced Pitch Class Profile' by Kyogu Lee\nThis script computes 12 dimensional chromagram for chord detection\n@author ORCHISAMA\n\"\"\"\n\nfrom __future__ import division\nfrom scipy.signal import hamming\nfrom scipy.fftpack import fft\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n\ndef nearestPow2(inp):\n power = np.ceil(np.log2(inp))\n return 2**power\n\n\n\"\"\"Function to calculcate Harmonic Power Spectrum from DFT\"\"\"\n\n\ndef HPS(dft, M):\n\n hps_len = int(np.ceil(np.size(dft) / (2**M)))\n hps = np.ones(hps_len)\n for n in range(hps_len):\n for m in range(M + 1):\n hps[n] *= np.absolute(dft[(2**m) * n])\n return hps\n\n\n\"\"\"Function to compute CQT using sparse matrix multiplication, Brown and Puckette 1992- fast\"\"\"\n\n\ndef CQT_fast(x, fs, bins, fmin, fmax, M):\n\n threshold = 0.0054 # for Hamming window\n K = int(bins * np.ceil(np.log2(fmax / fmin)))\n Q = 1 / (2 ** (1 / bins) - 1)\n nfft = np.int32(nearestPow2(np.ceil(Q * fs / fmin)))\n tempKernel = np.zeros(nfft, dtype=np.complex)\n specKernel = np.zeros(nfft, dtype=np.complex)\n sparKernel = []\n\n # create sparse Kernel\n for k in range(K - 1, -1, -1):\n fk = (2 ** (k / bins)) * fmin\n N = np.int32(np.round((Q * fs) / fk))\n tempKernel[:N] = hamming(N) / N * np.exp(-2 * np.pi * 1j * Q * np.arange(N) / N)\n specKernel = fft(tempKernel)\n specKernel[np.where(np.abs(specKernel) <= threshold)] = 0\n if k == K - 1:\n sparKernel = specKernel\n else:\n sparKernel = np.vstack((specKernel, sparKernel))\n\n sparKernel = np.transpose(np.conjugate(sparKernel)) / nfft\n ft = fft(x, nfft)\n cqt = np.dot(ft, sparKernel)\n ft = fft(x, nfft * (2**M))\n # calculate harmonic power spectrum\n # harm_pow = HPS(ft,M)\n # cqt = np.dot(harm_pow, sparKernel)\n return cqt\n\n\n\"\"\"Function to compute constant Q Transform, Judith Brown, 1991 - slow\"\"\"\n\n\ndef CQT_slow(x, fs, bins, fmin, fmax):\n\n K = int(bins * np.ceil(np.log2(fmax / fmin)))\n Q = 1 / (2 ** (1 / bins) - 1)\n cqt = np.zeros(K, dtype=np.complex)\n\n for k in range(K):\n fk = (2 ** (k / bins)) * fmin\n N = int(np.round(Q * fs / fk))\n arr = -2 * np.pi * 1j * Q * np.arange(N) / N\n cqt[k] = np.dot(x[:N], np.transpose(hamming(N) * np.exp(arr))) / N\n return cqt\n\n\n\"\"\"Function to compute Pitch Class Profile from constant Q transform\"\"\"\n\n\ndef PCP(cqt, bins, M):\n CH = np.zeros(bins)\n for b in range(bins):\n CH[b] = np.sum(cqt[b + (np.arange(M) * bins)])\n return CH\n\n\ndef compute_chroma(x, fs):\n\n fmin = 96\n fmax = 5250\n bins = 12\n M = 3\n nOctave = np.int32(np.ceil(np.log2(fmax / fmin)))\n CH = np.zeros(bins)\n # Compute constant Q transform\n cqt_fast = CQT_fast(x, fs, bins, fmin, fmax, M)\n # get Pitch Class Profile\n CH = PCP(np.absolute(cqt_fast), bins, nOctave)\n return CH\n" }, { "path": "create_templates.py", "content": "\"\"\"\nAlgorithm based on the paper 'Automatic Chord Recognition from\nAudio Using Enhanced Pitch Class Profile' by Kyogu Lee\nThis script computes 12 dimensional chromagram for chord detection\n@author ORCHISAMA DAS\n\"\"\"\n\n\"\"\"Create pitch profile template for 12 major and 12 minor chords and save them in a json file\nGmajor template = [1,0,0,0,1,0,0,1,0,0,0,0] - needs to be run just once\"\"\"\n\nimport json\n\ntemplate = dict()\nmajor = [\"G\", \"G#\", \"A\", \"A#\", \"B\", \"C\", \"C#\", \"D\", \"D#\", \"E\", \"F\", \"F#\"]\nminor = [\"Gm\", \"G#m\", \"Am\", \"A#m\", \"Bm\", \"Cm\", \"C#m\", \"Dm\", \"D#m\", \"Em\", \"Fm\", \"F#m\"]\noffset = 0\nnum_chords = len(major)\n\n# initialise lists with zeros\nfor chord in range(num_chords):\n template[major[chord]] = list()\n template[minor[chord]] = list()\n for note in range(num_chords):\n template[major[chord]].append(0)\n template[minor[chord]].append(0)\n\nfor chord in range(num_chords):\n for note in range(num_chords):\n if note == 0 or note == 7:\n template[major[chord]][(note + offset) % num_chords] = 1\n template[minor[chord]][(note + offset) % num_chords] = 1\n elif note == 4:\n template[major[chord]][(note + offset) % num_chords] = 1\n elif note == 3:\n template[minor[chord]][(note + offset) % num_chords] = 1\n offset += 1\n\n# debugging\nfor key, value in template.items():\n print(key, value)\n\n# save as JSON file\nwith open(\"chord_templates.json\", \"w\") as fp:\n json.dump(template, fp, sort_keys=False)\n print(\"Saved succesfully to JSON file\")\n" }, { "path": "data/chord_templates.json", "content": "{\"A#m\": [0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0], \"C#m\": [0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0], \"A#\": [0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0], \"Dm\": [0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 0], \"C#\": [0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0], \"Bm\": [0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1], \"G#\": [0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0], \"Fm\": [0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0], \"A\": [0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0], \"C\": [1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0], \"B\": [0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1], \"E\": [0, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0], \"D\": [0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1], \"G\": [1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0], \"F\": [0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0], \"G#m\": [0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0], \"Em\": [1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0], \"D#m\": [0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1], \"Cm\": [1, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0], \"Am\": [0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0], \"D#\": [1, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0], \"F#\": [0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1], \"Gm\": [1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0], \"F#m\": [0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1]}" }, { "path": "data/test_chords/readme.txt", "content": "piano chords downloaded from http://ibeat.org/piano-chords-free/. Give them credit under the Creative Commons License." }, { "path": "hmm.py", "content": "\"\"\"Automatic chord recogniton with HMM, as suggested by Juan P. Bello in\n'A mid level representation for harmonic content in music signals'\n@author ORCHISAMA DAS, 2016\"\"\"\n\nfrom __future__ import division\nfrom chromagram import compute_chroma\nimport os\nimport numpy as np\n\n\n\"\"\"calculates multivariate gaussian matrix from mean and covariance matrices\"\"\"\n\n\ndef multivariate_gaussian(x, meu, cov):\n\n det = np.linalg.det(cov)\n val = np.exp(-0.5 * np.dot(np.dot((x - meu).T, np.linalg.inv(cov)), (x - meu)))\n try:\n val /= np.sqrt(((2 * np.pi) ** 12) * det)\n except:\n print(\"Matrix is not positive, semi-definite\")\n if np.isnan(val):\n val = np.finfo(float).eps\n return val\n\n\n\"\"\"initialize the emission, transition and initialisation matrices for HMM in chord recognition\nPI - initialisation matrix, #A - transition matrix, #B - observation matrix\"\"\"\n\n\ndef initialize(chroma, templates, chords, nested_cof):\n\n \"\"\"initialising PI with equal probabilities\"\"\"\n num_chords = len(chords)\n PI = np.ones(num_chords) / num_chords\n\n \"\"\"initialising A based on nested circle of fifths\"\"\"\n eps = 0.01\n A = np.empty((num_chords, num_chords))\n for chord in chords:\n ind = nested_cof.index(chord)\n t = ind\n for i in range(num_chords):\n if t >= num_chords:\n t = t % num_chords\n A[ind][t] = (abs(num_chords // 2 - i) + eps) / (\n num_chords**2 + num_chords * eps\n )\n t += 1\n\n \"\"\"initialising based on tonic triads - Mean matrix; Tonic with dominant - 0.8,\n tonic with mediant 0.6 and mediant-dominant 0.8, non-triad diagonal elements \n with 0.2 - covariance matrix\"\"\"\n\n nFrames = np.shape(chroma)[1]\n B = np.zeros((num_chords, nFrames))\n meu_mat = np.zeros((num_chords, num_chords // 2))\n cov_mat = np.zeros((num_chords, num_chords // 2, num_chords // 2))\n meu_mat = np.array(templates)\n offset = 0\n\n for i in range(num_chords):\n if i == num_chords // 2:\n offset = 0\n tonic = offset\n if i < num_chords // 2:\n mediant = (tonic + 4) % (num_chords // 2)\n else:\n mediant = (tonic + 3) % (num_chords // 2)\n dominant = (tonic + 7) % (num_chords // 2)\n\n # weighted diagonal\n cov_mat[i, tonic, tonic] = 0.8\n cov_mat[i, mediant, mediant] = 0.6\n cov_mat[i, dominant, dominant] = 0.8\n\n # off-diagonal - matrix not positive semidefinite, hence determinant is negative\n # for n in [tonic,mediant,dominant]:\n # for m in [tonic, mediant, dominant]:\n # if (n is tonic and m is mediant) or (n is mediant and m is tonic):\n # cov_mat[i,n,m] = 0.6\n # else:\n # cov_mat[i,n,m] = 0.8\n\n # filling non zero diagonals\n for j in range(num_chords // 2):\n if cov_mat[i, j, j] == 0:\n cov_mat[i, j, j] = 0.2\n offset += 1\n\n \"\"\"observation matrix B is a multivariate Gaussian calculated from mean vector and \n covariance matrix\"\"\"\n\n for m in range(nFrames):\n for n in range(num_chords):\n B[n, m] = multivariate_gaussian(\n chroma[:, m], meu_mat[n, :], cov_mat[n, :, :]\n )\n\n return (PI, A, B)\n\n\n\"\"\"Viterbi algorithm to find Path with highest probability - dynamic programming\"\"\"\n\n\ndef viterbi(PI, A, B):\n (nrow, ncol) = np.shape(B)\n path = np.zeros((nrow, ncol))\n states = np.zeros((nrow, ncol))\n path[:, 0] = PI * B[:, 0]\n\n for i in range(1, ncol):\n for j in range(nrow):\n s = [(path[k, i - 1] * A[k, j] * B[j, i], k) for k in range(nrow)]\n (prob, state) = max(s)\n path[j, i] = prob\n states[j, i - 1] = state\n\n return (path, states)\n" }, { "path": "main.py", "content": "import numpy as np\nimport os, sys, getopt\nimport matplotlib.pyplot as plt\nfrom scipy.io.wavfile import read\nimport json\nfrom chromagram import compute_chroma\nimport hmm as hmm\n\n\ndef get_templates(chords):\n \"\"\"read from JSON file to get chord templates\"\"\"\n with open(\"data/chord_templates.json\", \"r\") as fp:\n templates_json = json.load(fp)\n templates = []\n\n for chord in chords:\n if chord == \"N\":\n continue\n templates.append(templates_json[chord])\n\n return templates\n\n\ndef get_nested_circle_of_fifths():\n chords = [\n \"N\",\n \"G\",\n \"G#\",\n \"A\",\n \"A#\",\n \"B\",\n \"C\",\n \"C#\",\n \"D\",\n \"D#\",\n \"E\",\n \"F\",\n \"F#\",\n \"Gm\",\n \"G#m\",\n \"Am\",\n \"A#m\",\n \"Bm\",\n \"Cm\",\n \"C#m\",\n \"Dm\",\n \"D#m\",\n \"Em\",\n \"Fm\",\n \"F#m\",\n ]\n nested_cof = [\n \"G\",\n \"Bm\",\n \"D\",\n \"F#m\",\n \"A\",\n \"C#m\",\n \"E\",\n \"G#m\",\n \"B\",\n \"D#m\",\n \"F#\",\n \"A#m\",\n \"C#\",\n \"Fm\",\n \"G#\",\n \"Cm\",\n \"D#\",\n \"Gm\",\n \"A#\",\n \"Dm\",\n \"F\",\n \"Am\",\n \"C\",\n \"Em\",\n ]\n return chords, nested_cof\n\n\ndef find_chords(\n x: np.ndarray,\n fs: int,\n templates: list,\n chords: list,\n nested_cof: list = None,\n method: str = None,\n plot: bool = False,\n):\n \"\"\"\n Given a mono audio signal x, and its sampling frequency, fs,\n find chords in it using 'method'\n Args:\n x : mono audio signal\n fs : sampling frequency (Hz)\n templates: dictionary of chord templates\n chords: list of chords to search over\n nested_cof: nested circle of fifth chords\n method: template matching or HMM\n plot: if results should be plotted\n \"\"\"\n\n # framing audio, window length = 8192, hop size = 1024 and computing PCP\n nfft = 8192\n hop_size = 1024\n nFrames = int(np.round(len(x) / (nfft - hop_size)))\n # zero padding to make signal length long enough to have nFrames\n x = np.append(x, np.zeros(nfft))\n xFrame = np.empty((nfft, nFrames))\n start = 0\n num_chords = len(templates)\n chroma = np.empty((num_chords // 2, nFrames))\n id_chord = np.zeros(nFrames, dtype=\"int32\")\n timestamp = np.zeros(nFrames)\n max_cor = np.zeros(nFrames)\n\n # step 1. compute chromagram\n for n in range(nFrames):\n xFrame[:, n] = x[start : start + nfft]\n start = start + nfft - hop_size\n timestamp[n] = n * (nfft - hop_size) / fs\n chroma[:, n] = compute_chroma(xFrame[:, n], fs)\n\n if method == \"match_template\":\n # correlate 12D chroma vector with each of\n # 24 major and minor chords\n for n in range(nFrames):\n cor_vec = np.zeros(num_chords)\n for ni in range(num_chords):\n cor_vec[ni] = np.correlate(chroma[:, n], np.array(templates[ni]))\n max_cor[n] = np.max(cor_vec)\n id_chord[n] = np.argmax(cor_vec) + 1\n\n # if max_cor[n] < threshold, then no chord is played\n # might need to change threshold value\n id_chord[np.where(max_cor < 0.8 * np.max(max_cor))] = 0\n final_chords = [chords[cid] for cid in id_chord]\n\n elif method == \"hmm\":\n # get max probability path from Viterbi algorithm\n (PI, A, B) = hmm.initialize(chroma, templates, chords, nested_cof)\n (path, states) = hmm.viterbi(PI, A, B)\n\n # normalize path\n for i in range(nFrames):\n path[:, i] /= sum(path[:, i])\n\n # choose most likely chord - with max value in 'path'\n final_chords = []\n indices = np.argmax(path, axis=0)\n final_states = np.zeros(nFrames)\n\n # find no chord zone\n set_zero = np.where(np.max(path, axis=0) < 0.3 * np.max(path))[0]\n if np.size(set_zero) > 0:\n indices[set_zero] = -1\n\n # identify chords\n for i in range(nFrames):\n if indices[i] == -1:\n final_chords.append(\"NC\")\n else:\n final_states[i] = states[indices[i], i]\n final_chords.append(chords[int(final_states[i])])\n\n if plot:\n plt.figure()\n if method == \"match_template\":\n plt.yticks(np.arange(num_chords + 1), chords)\n plt.plot(timestamp, id_chord, marker=\"o\")\n\n else:\n plt.yticks(np.arange(num_chords), chords)\n plt.plot(timestamp, np.int32(final_states), marker=\"o\")\n\n plt.xlabel(\"Time in seconds\")\n plt.ylabel(\"Chords\")\n plt.title(\"Identified chords\")\n plt.grid(True)\n plt.show()\n\n return timestamp, final_chords\n\n\ndef main(argv):\n input_file = \"\"\n method = \"\"\n plot = False\n has_method = False\n try:\n opts, args = getopt.getopt(argv, \"hi:m:p:\", [\"ifile=\", \"method=\", \"plot=\"])\n except getopt.GetoptError:\n print(\"main.py -i