The research team of Tsinghua University has developed a mobile sound source simulation platform called SonicSim, aiming to solve the problem of data scarcity in mobile sound source scenarios in the field of speech processing. The platform is built on Habitat-sim and can highly restore the real acoustic environment and provide high-quality data for the training and evaluation of speech separation and enhanced models. Most existing data sets are based on static sound sources and are difficult to meet actual needs. However, the scale of real-recorded data sets is limited and costly, while the synthetic data sets lack authenticity. The SonicSim platform effectively solves these problems and builds a large multi-scenario mobile sound source dataset SonicSet.
This platform can simulate a variety of complex acoustic environments, including obstacle occlusion, room geometry, and the impact of different materials on sound, and supports user-defined scene parameters. The SonicSet dataset utilizes data from LibriSpeech, Freesound Dataset50k, and Free Music Archive, as well as real scenes from the Matterport3D dataset, and contains rich voice, ambient noise and music noise data. Its construction process is highly automated, ensuring the authenticity and diversity of data. Experimental results show that the model trained on the SonicSet dataset performs better on the real dataset, verifying the effectiveness of the SonicSim platform. The release of the SonicSim platform and SonicSet dataset has brought new breakthroughs to the field of speech processing, and will further promote the application of speech processing technology in complex environments in the future, but its authenticity is still limited by the details of 3D scene modeling. Paper address: https://arxiv.org/pdf/2410.01481
The emergence of the SonicSim platform provides new ideas for data acquisition in the field of speech processing, and also highlights the important role of simulation technology in solving practical problems. In the future, with the continuous development of technology, I believe that similar simulation platforms will play a role in more fields and promote the progress of artificial intelligence technology.