Target sound information extraction: Speech and audio processing with neural networks conditioned on target clues

抄録

This paper overviews neural target sound information extraction (TSIE), which consists of extracting the desired information about a sound source in an observed sound mixture given clues about the target source. TSIE is a general framework, which covers various applications, such as target speech/sound extraction (TSE), personalized voice activity detection (PVAD), target speaker automatic speech recognition (TS-ASR), etc. We formalize the ideas of TSIE and show how it can be implemented through various examples such as TSE, PVAD, and TS-ASR. We conclude the paper with a discussion of potential future research directions.