Music Fingerprinting using Locality Sensitive Hashing

Music Fingerprinting using Locality Sensitive Hashing#

import librosa
import os.path
import numpy as np

from mirdotcom import mirdotcom

mirdotcom.init()

This notebook shows a simple system for performing retrieval of musical tracks using LSH.

Select training data:

training_dir = mirdotcom.AUDIO_DIRECTORY + "drum_samples/train/"
training_files = [os.path.join(training_dir, f) for f in os.listdir(training_dir)]

Define a hash function:

def hash_func(vecs, projections):
    bools = np.dot(vecs, projections.T) > 0
    return [bool2int(bool_vec) for bool_vec in bools]

def bool2int(x):
    y = 0
    for i, j in enumerate(x):
        if j:
            y += 1 << i
    return y

bool2int([False, True, False, True])

10

X = np.random.randn(10, 100)
P = np.random.randn(3, 100)
hash_func(X, P)

[0, 6, 1, 5, 0, 1, 4, 6, 4, 5]

Create three LSH structures: Table, LSH, and MusicSearch:#

class Table:
    def __init__(self, hash_size, dim):
        self.table = dict()
        self.hash_size = hash_size
        self.projections = np.random.randn(self.hash_size, dim)

    def add(self, vecs, label):
        entry = {"label": label}
        hashes = hash_func(vecs, self.projections)
        for h in hashes:
            if h in self.table.keys():
                self.table[h].append(entry)
            else:
                self.table[h] = [entry]

    def query(self, vecs):
        hashes = hash_func(vecs, self.projections)
        results = list()
        for h in hashes:
            if h in self.table.keys():
                results.extend(self.table[h])
        return results

class LSH:
    def __init__(self, dim):
        self.num_tables = 4
        self.hash_size = 8
        self.tables = list()
        for i in range(self.num_tables):
            self.tables.append(Table(self.hash_size, dim))

    def add(self, vecs, label):
        for table in self.tables:
            table.add(vecs, label)

    def query(self, vecs):
        results = list()
        for table in self.tables:
            results.extend(table.query(vecs))
        return results

    def describe(self):
        for table in self.tables:
            print(table.table)

class MusicSearch:
    def __init__(self, training_files):
        self.frame_size = 4096
        self.hop_size = 4000
        self.fv_size = 12
        self.lsh = LSH(self.fv_size)
        self.training_files = training_files
        self.num_features_in_file = dict()
        for f in self.training_files:
            self.num_features_in_file[f] = 0

    def train(self):
        for filepath in self.training_files:
            x, fs = librosa.load(filepath)
            features = librosa.feature.chroma_stft(
                y=x, sr=fs, n_fft=self.frame_size, hop_length=self.hop_size
            ).T
            self.lsh.add(features, filepath)
            self.num_features_in_file[filepath] += len(features)

    def query(self, filepath):
        x, fs = librosa.load(filepath)
        features = librosa.feature.chroma_stft(
            y=x, sr=fs, n_fft=self.frame_size, hop_length=self.hop_size
        ).T
        results = self.lsh.query(features)
        print("num results", len(results))

        counts = dict()
        for r in results:
            if r["label"] in counts.keys():
                counts[r["label"]] += 1
            else:
                counts[r["label"]] = 1
        for k in counts:
            counts[k] = float(counts[k]) / self.num_features_in_file[k]
        return counts

Train:

ms = MusicSearch(training_files)
ms.train()

Test:

test_file = mirdotcom.get_audio("drum_samples/test/kick_00.mp3")
results = ms.query(test_file)

num results 373

Display the results:

for r in sorted(results, key=results.get, reverse=True):
    print(r, results[r])

assets/audio/drum_samples/train/kick_09.mp3 6.0
assets/audio/drum_samples/train/snare_06.mp3 6.0
assets/audio/drum_samples/train/kick_02.mp3 4.666666666666667
assets/audio/drum_samples/train/kick_03.mp3 4.375
assets/audio/drum_samples/train/kick_01.mp3 4.0
assets/audio/drum_samples/train/snare_07.mp3 4.0
assets/audio/drum_samples/train/kick_10.mp3 3.7777777777777777
assets/audio/drum_samples/train/snare_03.mp3 3.75
assets/audio/drum_samples/train/kick_06.mp3 3.625
assets/audio/drum_samples/train/snare_01.mp3 3.25
assets/audio/drum_samples/train/snare_10.mp3 2.75
assets/audio/drum_samples/train/kick_08.mp3 2.5
assets/audio/drum_samples/train/kick_05.mp3 2.111111111111111
assets/audio/drum_samples/train/snare_04.mp3 1.8
assets/audio/drum_samples/train/kick_04.mp3 1.75
assets/audio/drum_samples/train/snare_05.mp3 1.75
assets/audio/drum_samples/train/snare_08.mp3 1.75
assets/audio/drum_samples/train/kick_07.mp3 1.6666666666666667
assets/audio/drum_samples/train/snare_02.mp3 1.0
assets/audio/drum_samples/train/snare_09.mp3 0.25
Contents