How do I read in wav files in .gz?

Question

I am learning machine learning and data analysis on wav files. I know if I have wav files directly I can do something like this to read in the data

import librosa

mono, fs = librosa.load('./small_data/time_series_audio.wav', sr = 44100)

Now I'm given a gz-file "music_feature_extraction_test.tar.gz" I'm not sure what to do now.

I tried:

with gzip.open('music_train.tar.gz', 'rb') as f:
    for files in f : 
        mono, fs = librosa.load(files, sr = 44100)

but it gives me:

TypeError: lstat() argument 1 must be encoded string without null bytes, not str

Can anyone help me out?

Answer 1

There are several things going on:

• The file you are given is a gzipped-compressed tarball. Take a look at the tarfile module , it can read gzip-compressed files directly. You'll get an iterator over it's members, each of which is an individual file.
• AFAIKS librosa can't read from an in-memory buffer so you have to unpack the tar-members to temporary files. The tempfile - module is your friend here, a NamedTemporaryFile will provide you with a self-deleting file that you can uncompress to and provide to librosa .

You probably want to implement this as a simple generator function that takes the tarfile-name as it's input, iterates over it's members and yield s what librosa.load() provides you. That way everything gets cleaned up automatically.

The basic loop would therefore be

1. Open the tarball using the tarfile -module. For each member
2. Get a new temporary file using NamedTemporaryFile . Copy the content of the tarball-member to that file. You may want to use shutil.copyfileobj to avoid reading the entire wav-file into memory before writing it to disk.
3. The NamedTemporaryFile has a filename-attribute. Pass that to librosa.open .
4. yield the return value of librosa.open to the caller.

Answer 2

You can use PySoundFile to read from the compressed file. https://pysoundfile.readthedocs.io/en/0.9.0/#virtual-io

import soundfile

with gzip.open('music_train.tar.gz', 'rb') as gz_f:
    for file in gz_f : 
        fs, mono = soundfile.read(file, samplerate=44100)

Maybe you should also check if you need to resample the data before processing it with librosa: https://librosa.github.io/librosa/ioformats.html#read-specific-formats