Neural-Network-Based Spectrum Processing for Speech Recognition and Speaker Verification

摘要

In this paper, neural networks are applied as a feature extractors for a speech recognition system and a speaker verification system. A long-temporal features with delta coefficients, mean and variance normalization are applied when a neural-network-based feature extraction is trained together with a neural-network-based voice activity detector and with a neural-network-based acoustic model for speech recognition. In speaker verification, the acoustic model is replaced with a score computation. The performance of our speech recognition system was evaluated on the British English speech corpus WSJCAM0 and the performance of our speech verification system was evaluated on our Czech speech corpus.