A voice-operated transmit (VOX) algorithm for input sinal with low SNR

This paper presents a voice-operated transmit (VOX) algorithm for input signal with low signal-noise rate (SNR). The algorithm integrates a noise reduction algorithm and a voice-active detection (VAD) algorithm. A perfect reconstruction filter bank decomposes full-band time series into multiple Mel-scaled sub-bands. A voice-active detection is performed in each sub-band, and the results are used to dynamically adjust the background noise gain applied to each sub-band. Upon reconstruction of the full-band time series, the background noise level is reduced relative to the speech. For many applications, 15 to 25 dB of noise reduction can be achieved in real-world settings. The method will be demonstrated using examples of speech recorded in excavator. As demonstrated, the method can dramatically enhance the quality of speech.