Attention: This operator is going to be merged into the TensorFlow-IO project (PR #815). Please use that version instead - I only keep this repository to have a minimal real-life example of a custom TF operator.
An operator to decode MP3 data from in-memory data similar to tf.audio.decode_wav. This is handy to decode large numbers of small MP3 files, e.g. when reading from a TFRecord file. It uses @lieff's awesome minimp3 library.
Warning: Not tested with stereo files yet.
Only local installation supported so far.
cd tf-decode-mp3
make
pip install -e .
import tensorflow as tf
from tf_decode_mp3 import decode_mp3
ds = (tf.data.Dataset
.list_files("*.mp3")
.map(tf.io.read_file)
.map(decode_mp3))
for samples, sample_rate in ds:
print(samples.shape, sample_rate)
The decode_mp3 function has two additional parameters:
desired_channels: Fix how many channels are returned. If the file contains more, only the first is returned. If the file contains less, they are duplicated.desired_samples: Fix how many samples are returned per channel. If the file contains more, the surplus is thrown away. If the file contains less, they are padded on the right with 0s.