Improving Reverberant Speech Training Using Diffuse Acoustic Simulation
This work addresses the challenge of data scarcity and realism in training for speech recognition and keyword spotting, particularly in reverberant environments, offering a practical solution for improving model performance.
The paper tackles the problem of generating realistic training data for speech-related machine learning tasks by developing an efficient geometric acoustic simulation method that models occlusion, specular, and diffuse reflections in complex environments. The result shows significant performance improvements on real test sets, with a 1.58% gain in far-field speech recognition and a 21% gain in keyword spotting, without fine-tuning using real impulse responses.
We present an efficient and realistic geometric acoustic simulation approach for generating and augmenting training data in speech-related machine learning tasks. Our physically-based acoustic simulation method is capable of modeling occlusion, specular and diffuse reflections of sound in complicated acoustic environments, whereas the classical image method can only model specular reflections in simple room settings. We show that by using our synthetic training data, the same neural networks gain significant performance improvement on real test sets in far-field speech recognition by 1.58% and keyword spotting by 21%, without fine-tuning using real impulse responses.