Noise robust spontaneous speech recognition using multi-space GMM

In this paper, we propose a new approach using a multi-space Gaussian mixture model (GMM) for a large-scale spontaneous speech recognition system that is robust to the acoustic environmental noise. Current speech recognition systems based on a hidden Markov model (HMM) perform well in matched conditions, but their performance is degraded by mismatch conditions, such as mobile environments with diverse additive noise. In the case of mobile voice search services, the real noise environment is reflected in rich speech log data, and using speech logs, performance improvement is achieved in the growing matched condition. However, because most of this speech data is short with a limited pattern, when it is used for large-scale spontaneous speech recognition tasks like voice SMS, the performance improvement is limited and degradation is even observed in a quiet environment. Therefore, this paper proposes a new approach which, using rich voice search speech data, constructs a multi- Acoustic space GMM with distributions of speech corrupted by diverse environment noise and reflects these statistics in an acoustic model for a speech recognition system with a distinct domain like dictation speech. The evaluation results obtained from the voice SMS task show that the proposed method provides meaningful improvements over conventional adaptive training methods to handle multi-style training data. Copyright