Exploring the robustness of features and enhancement on speech recognition systems in highly-reverberant real environments
This work addresses speech recognition performance in challenging real-world acoustic conditions, but it is incremental as it compares existing enhancement and feature methods.
This paper evaluated the robustness of a DNN-HMM speech recognition system in highly-reverberant real environments, finding that using Weighted Prediction Error (WPE) enhancement with locally-normalized filter bank (LNFB) features reduced word error rates by 3% and 20% on average compared to other methods under reverberated training conditions.
This paper evaluates the robustness of a DNN-HMM-based speech recognition system in highly-reverberant real environments using the HRRE database. The performance of locally-normalized filter bank (LNFB) and Mel filter bank (MelFB) features in combination with Non-negative Matrix Factorization (NMF), Suppression of Slowly-varying components and the Falling edge (SSF) and Weighted Prediction Error (WPE) enhancement methods are discussed and evaluated. Two training conditions were considered: clean and reverberated (Reverb). With Reverb training the use of WPE and LNFB provides WERs that are 3% and 20% lower in average than SSF and NMF, respectively. WPE and MelFB provides WERs that are 11% and 24% lower in average than SSF and NMF, respectively. With clean training, which represents a significant mismatch between testing and training conditions, LNFB features clearly outperform MelFB features. The results show that different types of training, parametrization, and enhancement techniques may work better for a specific combination of speaker-microphone distance and reverberation time. This suggests that there could be some degree of complementarity between systems trained with different enhancement and parametrization methods.