Voice activity detection with quasi-quadrature filters and GMM decomposition for speech and noise

Ryszard Makowski, Robert Hossa

Applied Acoustics · 2020 · 11 citations · 19 references

DOIFull text

Open access

Concepts

Abstract

One of basic algorithms employed in technologies such as automatic speech recognition (ASR) systems is voice activity detection (VAD). Speech contains many pauses, whose interpretation might lead to recognition errors. Scientific literature provides numerous VAD algorithms, though many of them have substantial memory and/or calculation time requirements. On the other hand, the efficacy of those with smaller requirements is usually unsatisfactory. This paper proposes a modification to a single frequency filtering based algorithm known from literature, changing the methods of determining envelopes and the detection threshold. The purpose of these modifications was to reduce the calculation time and memory requirements without losing the efficiency of the algorithm. Also, a completely new algorithm of determining the detection threshold, using the approximation of hypothesis probability distribution was developed. The obtained results are satisfactory.

References

19