Speaker independent speech recognition system based on phoneme identification

Speaker independent speech recognition is important for successful development of speech recognizers in most real world applications. While speaker dependent speech recognizers have achieved close to 100% accuracy, the speaker independent speech recognition systems have poor accuracy not exceeding 75%. In this paper we describe a two-module speaker independent speech recognition system for all-Indian English speech. The first module performs phoneme recognition using two-level neural networks. The second module executes word recognition from the string of phonemes employing Hidden Markov Model. The system was trained by Indian English speech consisting of 3000 words uttered by 200 speakers. The test samples comprised 1000 words spoken by a different set of 50 speakers. The recognition accuracy is found to be 94% which is well above the previous results.