This repository contains code associated with the paper titled "Artificially Synthesising Data for Audio Classification and Segmentation to Improve Speech and Music Detection in Radio Broadcast" accepted for publication in IEEE ICASSP 2021 [PDF].
@InProceedings{venkatesh2021artificially,
author = {Satvik Venkatesh and David Moffat and Alexis Kirke and G\"{o}zel Shakeri and Stephen Brewster and J\"{o}rg Fachner and Helen Odell-Miller and Alex Street and Nicolas Farina and Sube Banerjee and Eduardo Reck Miranda},
booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title = {{Artificially Synthesising Data for Audio Classification and Segmentation to Improve Speech and Music Detection in Radio Broadcast}},
year = {2021},
}Machine learning models for audio segmentation and music-speech detection are generally trained on proprietary audio, which cannot be shared. Labelling this data is time-consuming and expensive, which discourages new researchers in this field. In this project, we artificially synthesise data that resembles radio signals by replicating the workflow of a radio DJ.
The file data-synthesis.ipynb contains the code for artificially synthesising data. The synthesised data can be stored in your personal Google Drive. The file train-CRNN.ipynb contains the code to train a Convolutional Recurrent Neural Network on the synthesised data. The file detection-example.ipynb performs segmentation over any audio file using the pre-trained model.
A few synthetic examples are available in the Synthetic Radio Examples folder.
Please note that the pre-trained models are for non-commercial use only. It has been trained on some datasets that do not permit commercial use. The entire list of datasets used for training are MUSAN, GTZAN music-speech, GTZAN Genre collection, Scheirer & Slaney, Instrument Recognition in Musical Audio Signals, Singing Voice dataset, and LibriSpeech.
Want to try it without installing anything? Here is a Google Colab notebook.
The pre-trained models are under the Create Commons Attribution-NonCommercial-ShareAlike-3.0 Unported (CC BY-NC-SA 3.0) license, and the source code for the project is under the MIT license.