Evolutionary Algorithms and Speech Recognition

In this chapter, we present an approach for optimizing the front-end processing of ASR systems by using Genetic Algorithms (GAs). The front-end uses a multi-stream approach to incorporate, in addition to MFCCs, auditory-based phonetic distinctive cues. These features are combined in order to limit the impact of the speech signal degradations due to interfering noise. Some of many advantages of using GAs include the possibility to improve the robustness without modifying the recognition system models and without estimating environment parameters, such as the noise variance and/or stream weights. The co-existence, in two streams, of the two types of front-end parameters (MFCCs and distinctive cues) is also managed by the GA. The evaluation is carried out by using a noisy version of the TIMIT corpus.