Audio model for Speech/Music/SFX discrimination based on VGGish and Modelling of Experts' Decisions on assigning importances on radio objects

Introduction

This model is provided as supplementary material for the following paper:

@inproceedings{choudakisward2019dafx,
  title={MODELLING EXPERTS' DECISIONS ON ASSIGNING NARRATIVE IMPORTANCES OF
OBJECTS IN A RADIO DRAMA MIX},
  author={Chourdakis, E. T. and Ward, L. and Paradis, M. and Reiss, J.D. },
  year={2019},
  booktitle={Int. Conf. Digital Audio Effects},
  pages={[Under Review]},
}

Requirements

The model has been found to be working with the following versions

Module	Version
python3	3.6.7
tensorflow	1.12.0
keras	2.2.4

There are also several other requirements listed in requirements.txt

Installation

Please run the following:

git clone git@github.com:bbc/audio-dafx2019-automatic.git

cd audio-dafx2019-automatic

Download the .h5 and .json files from v004 to the same directory.

# (optional) 
virtualenv venv
source venv/bin/activate
# (/optional)

pip3 install -r requirements.txt

You also need to download vggish_input.py, mel_features.py, and vggish_params.py put it into the same directory.

curl -O https://raw.githubusercontent.com/tensorflow/models/master/research/audioset/vggish_input.py
curl -O https://raw.githubusercontent.com/tensorflow/models/master/research/audioset/mel_features.py
curl -O https://raw.githubusercontent.com/tensorflow/models/master/research/audioset/vggish_params.py

and make sure you read and agree with their LICENSES

head -n 14 vggish_input.py
head -n 14 mel_features.py
head -n 14 vggish_params.py

Classifying audio to speech/music/SFX

Suppose you have a bunch of track files you need to classify to either speech/music or sound effects. First, you need to make sure that they are mono tracks 22050Hz/16bit (we recommend using SoX):

sox  input.wav -b 16 output.wav rate 22050 remix 1-2

Suppose your files are in a path audio/, you can classify the files in that folder by running:

python3 classify.py --model music_speech_sfx_discriminator audio/

You will then have a file output.csv with the results:

cat output.csv

Happy classifying~~!

Assigning importances to a directory of radio track stems

Suppose you have a folder of stereo stems (sampling rate does not matter) stems/ containing .wav files. In order to assign an importance level ( 0 -- low importance, 3 -- essential importance) to each of those, download the models seen in the previous section (Installation) to the same folder as assign.py and then run:

python3 assign.py stems

This will generate a file output.csv with the list of filenames for each stem, as well features and assigned importances. For more options see

python3 assign.py --help

Training the Music/Speech/SFx classification model

Please see MusicSpeechSFxDiscrimination.ipynb

Contact

Questions or issues about the model should either be raised here or addressed to Emmanouil Theofanis Chourdakis <e.t.chourdakis__Aaa.t!__qmul.ac.uk>