InterSpeech 2021

A new approach to speech activity detection (SAD) is presented in this work. It allows us to reduce the complexity and computation demands, namely in services that process streaming speech, where a SAD module usually forms the first block of the data pipeline (e.g., in a platform for 24/7 broadcast transcription). Our approach utilizes x-vectors as input features so that, within the subsequent pipeline stages, these embedding instances can also directly be employed for speaker diarization and recognition. The x-vectors are extracted by feed-forward sequential memory network (FSMN), allowing for modeling long-time dependencies; they thus form an input into a computationally undemanding binary classifier, whose output is smoothed by a decoder. Evaluation is performed on the standardized QUT-NOISE-TIMIT dataset as well as on broadcast data with large portions of music and background noise. The former data allows for comparison with other existing approaches. The latter shows the performance in terms of word error rate (WER) and reduction in real-time factor (RTF) of the transcription process.