Improved speech inversion using general regression neural network.

Area of Science:

• Speech processing and machine learning
• Acoustic-to-articulatory inversion modeling

Background:

• Estimating articulatory features from acoustic speech signals is crucial for enhancing speech recognition.
• Existing state-of-the-art models like Deep Belief Networks (DBN) face challenges in nonlinear inversion mapping.

Purpose of the Study:

• To investigate and compare the performance of Deep Belief Networks (DBN) and General Regression Neural Networks (GRNN) for acoustic-to-articulatory inversion.
• To evaluate the effectiveness of GRNN-estimated articulatory features in improving speech recognition tasks.

Main Methods:

• Nonlinear acoustic-to-articulatory inversion was performed in the feature space using DBN and GRNN models.
• Experiments were conducted on the MOCHA-TIMIT and MNGU0 speech databases.
• Performance was evaluated based on root-mean-square error, correlation, phonetic classification, and phoneme recognition accuracy.

Main Results:

• GRNN demonstrated superior performance over DBN in speech inversion, achieving lower root-mean-square error and higher correlation.
• The combination of acoustic features and GRNN-estimated articulatory features resulted in state-of-the-art accuracy for broad class phonetic classification and phoneme recognition.
• This combined approach also required less computational power compared to DBN.

Conclusions:

• GRNN is a more effective model than DBN for acoustic-to-articulatory inversion.
• Integrating GRNN-derived articulatory features with acoustic data significantly enhances speech recognition performance and computational efficiency.