Expanded Speech Recognition to Include Foreign Accents

Current speech interfaces in many military applications may be adequate for native English speakers. However, recognition rates drop for users with foreign accents. Non-native speakers demonstrate large temporal (duration of each phoneme) and intra-phoneme (exact shape of speech signal) variations when pronouncing the same words. This problem is compounded by the presence of environmental noises from tanks, helicopters, etc. Intelligent Automation, Inc., with Prof. Richard Stern of Carnegie Mellon University (CMU) propose a novel integrated system to improve speech recognition performance for people with foreign accents. It is emphasized that this team has rich experience in understanding the speech characteristics in non-native speakers. Some adaptations are needed to update speaker models for non-native speakers. The proposed system uses state-of-the-art techniques and consists of the following five parts: 1) a speech enhancement module to eliminate background noise; 2) a speaker verification algorithm to identify the speaker; 3) a speaker adaptation algorithm that uses a small training data set to update the speakers' models in both the speaker verification and the speech recognition areas; 4) a speech recognition system to recognize speech; and 5) a speech coding module to compress speech for storage or transmission.