This is the code used to train a neural network with the goal of detecting of parts of a recording are speech or song.
The training data used is in forms of MP3 files (not included) and can be placed into the songs
or speech folders. Then you can run the ./split.sh scripts which splits the file into chunks
with the specified length. Currently only files with 44100 Hz sample rate are supported
Incomplete and untested list:
conda install -c apple tensorflow-deps
conda install tensorflow_io
conda install torchaudio
Run python3 train.py
This will produce a ./songornot_trained.pt file. If you run the script again the file will be replaced
As a sample you can run ./tests/inference_test.py <model> <audio file>.
Results will be similar to this:
/Users/matjaz/meeting.wav
Chunks: 1317
Total items: 1317
song (15:10): 00:00 - 15:10
speech (04:00): 15:10 - 19:10
song (03:05): 19:10 - 22:15
speech (18:30): 22:15 - 40:45
song (04:00): 40:45 - 44:45
speech (00:20): 44:45 - 45:05
song (00:15): 45:05 - 45:20
speech (08:05): 45:20 - 53:25
song (03:30): 53:25 - 56:55
speech (25:05): 56:55 - 82:00
song (01:55): 82:00 - 83:55
speech (20:40): 83:55 - 104:35
song (03:30): 104:35 - 108:05
The provided models are provided for demonstration purposes and are (C) 2023 BCC Media STI.
Model files listed here are licensed under CC BY-NC-ND 4.0 license:
- songornot_2s.pt
- songornot_5s.pt
Everything with the exception of the above list of files is released under the MIT License.