Have you ever heard or even use service like Shazam? Cool, right? No, we are not going to make something as magical as it ? But using chromaprint we can create audio fingerprint so that we can do music search by using a music sample.
Before we can use chromaprint python library, pyacoustid, we need to install chromaprint C library in our OS. The build instruction is in its repo. Or if you use Debian/Ubuntu, you can install it using apt (if there is any error, try to google up the error. Most of the time, it’s due to missing dependency):
sudo apt install ffmpeg acoustid-fingerprinter
Then you can create virtual environment (or not) and install the pip packages:
pip install pyacoustid
Now to generate a fingerprint from a file:
import acoustid
import chromaprint
duration, fp_encoded = acoustid.fingerprint_file('music.mp3')
fingerprint, version = chromaprint.decode_fingerprint(fp_encoded)
print(fingerprint)
The fingerprint will contain an array of 948 signed int that represent compact characteristic aka fingerprint of the audio file. It can also be visualized with the help of numpy and matplotlib:
pip install numpy matplotlib
import numpy as np
import matplotlib.pyplot as plt
...
fig = plt.figure()
bitmap = np.transpose(np.array([[b == '1' for b in list('{:32b}'.format(i & 0xffffffff))] for i in fingerprint]))
plt.imshow(bitmap)
Now, let say we have another music file and want to check the similarity with the previous file. One way to do this is by calculating similarity between the fingerprints:
pip install fuzzywuzzy[speedup]
from fuzzywuzzy import fuzz similarity = fuzz.ratio(sample_fingerprint, fingerprint) print(similarity)
The similarity will contains percentage of similarity between sample_fingerprint and fingerprint calculated using fuzzy algorithm.
I also made a simple program that uses all the codes above. The program calculates similarity between files in two directories, find the best match and also visualize the fingerprints.
That’s all. Thanks for reading ☕
Getting verified SSL information with Python (3.x) is very easy. Code examples for it are…
By default, Spring Data Couchbase implements single-bucket configuration. In this default implementation, all POJO (Plain…
Last year, Google released Firebase Auth Emulator as a new component in Firebase Emulator. In…
One of the authentication protocol that is supported by most of Google Cloud services is…
If you need to to add a spatial information querying in your application, PostGIS is…
Amazon Web Service Transcribe provides API to automatically convert an audio speech file (mp3/wav) into…
View Comments
Hello, this post is great. I'm trying to create a music service... I want to know how I can compare two big audio (radio fragments), then finding little similar fragments (songs), I need to split and extract the common fragments.
Thanks, beforehand!
Hi. I haven't done such thing before. If I should, maybe I'd try to find boundary between songs in the big audios then compare the signatures. You can train a machine learning model to detect the boundary https://pdfs.semanticscholar.org/0fee/81a489242eae8f0ef578321311481ffbda9f.pdf
Thanks. The text behind of link is very interesting, so I need to understand it. Can you explain me or recommend me something more, about that?
As far as I understand, the idea in the paper is to split audio into smaller chunks, manually labels them either as song or accompaniment (intro, filler between song .etc) and use them train a machine learning models (SVM). Unfortunately I can't explain more than that due to lack to knowledge & experience in audio processing.
If you prefer more practical recommendation, you can search in GitHub using "audio segmentation" keywords. This is the closest one I can find https://github.com/amsehili/audio-segmentation-by-classification-tutorial. It detects more classes than the paper, but you should be able to study it to get better idea.
Thanks, for your explanation... I'll researching.
Hi, Yohanes.
How are you?
I would like to know how to store the fingerprints in mongodb.
The fingerprint here is just array of signed int. So we can store it as it is in mongodb with the help of package such as `pymongo`. But the actual problem is how to search them again based on similarity.
I haven't done such thing before. But I would try to cluster the whole fingerprints (eg. using k-means) and store fingerprints along with their cluster id. Then to find the similarity, I can just (1) find the cluster id of the input, (2) query mongodb for all fingerprints with the same cluster id, (3) run text similarity on the results to find top matching fingerprints.
Hi. I would like to know, how I can do the first code on c++
//
duration, fp_encoded = acoustid.fingerprint_file('music.mp3')
fingerprint, version = chromaprint.decode_fingerprint(fp_encoded)
print(fingerprint)
Unfortunately, I haven't tried it in C++ and found no documentation. I've tried to skim pyaccoustid source code but it seems quite complex. If you really need to do it, you should try to trace it from there https://github.com/beetbox/pyacoustid
I tried typing in this code:
import acoustid
import chromaprint
duration, fp_encoded = acoustid.fingerprint_file('YTP.mp3')
fingerprint, version = chromaprint.decode_fingerprint(fp_encoded)
print(fingerprint)
but I get an error saying:
raise FingerprintGenerationError("audio could not be decoded")
acoustid.FingerprintGenerationError: audio could not be decoded
What could be the issue?
@Spencer I think it's dependency or audio file encoding issue. If you can share the mp3 file, I can try to process it my machine
Yes it seems that it is a dependency problem. :
freetype: no [The C/C++ header for freetype
(freetype2\ft2build.h) could not be found. You may
need to install the development package.]
png: no [The C/C++ header for png (png.h) could not be
found. You may need to install the development
package.]
However when I download both freetype and png it still cannot find them on my system.
for example I have done
PS C:\Users\sptzk> pip install pypng
Requirement already satisfied: pypng in c:......
and also tried to download from developers site.
is there a way to make it recognize that I do infact have it?
PS(truly appreciate your prompt help!!)
I
hatedon't use wind*ws for development, so I can't be certain. But based on the error message, it is still missing the libraries (freetype2\ft2build.handpng.h). So I think you need to install or copy them & verify their existance.Having python package installed doesn't guarantee the actual library installed. Because usually python packages only act as wrapper of the actual libraries or binaries.