A Front-End Solution to VAD, SNR, and AGC for Speech Recognition

In this paper VAD, SNR and AGC, as main complications of a front-end signal processor in automatic speech recognition are discussed with their structure and realization, including experimental evaluation on the module global-SNR is given. The front-end has five advantages: i) propose a novel scheme to long-term power estimation based on accumulate average; ii) elegant policy for VAD based on pitch detection; iii) relative highly accurate SNR estimation; iv) AGC based on soundcard when silence between utterances; v) robust performance even in very noisy acoustical circumstance.