Music and speech detection provides us valuable information regarding the nature of content in broadcast audio. It helps detect acoustic regions that contain speech, voice over music, only music, or silence. In recent years, there have been developments in machine learning algorithms to accomplish this task. However, broadcast audio is generally well-mixed and copyrighted, which makes it challenging to share across research groups. In this study, we address the challenges encountered in automatically synthesising data that resembles a radio broadcast. Firstly, we compare state-of-the-art neural network architectures such as CNN, GRU, LSTM, TCN, and CRNN. Later, we investigate how audio ducking of background music impacts the precision and recall of the machine learning algorithm. Thirdly, we examine how the quantity of synthetic training data impacts the results. Finally, we evaluate the effectiveness of synthesised, real-world, and combined approaches for training models, to understand if the synthetic data presents any additional value. Amongst the network architectures, CRNN was the best performing network. Results also show that the minimum level of audio ducking preferred by the machine learning algorithm was similar to that of human listeners. After testing our model on in-house and public datasets, we observe that our proposed synthesis technique outperforms real-world data in some cases and serves as a promising alternative.
This is an open access article distributed under the Creative Commons Attribution License
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited