Speech Recognition Using Bottleneck Features and Noise Elimination

The patent changes the parameter representation from traditional GMM-HMM features to bottleneck features extracted through deep neural networks. This parameter transformation enables more accurate HMM-state level information extraction by learning optimal feature representations that capture essential speech characteristics while reducing dimensionality and noise.

The patent substitutes the traditional GMM-HMM algorithmic approach with a deep neural network based bottleneck feature extraction system. This replacement transitions from conventional statistical modeling to learned representations, achieving superior accuracy in obtaining HMM-state level information while maintaining system reliability.

The patent extracts and removes background noise from training data through dedicated noise elimination processing. By separating and removing the harmful noise components while preserving the useful speech signals, the system maintains diversity in training data sources while ensuring high accuracy in HMM-state level information extraction.

The patent converts the presence of background noise in training data from a harmful factor into a beneficial preprocessing opportunity. By intentionally processing noisy data through noise elimination algorithms, the system transforms challenging real-world conditions into refined training samples that improve both model robustness and extraction accuracy.