Proportional-integral-derivative control of automatic speech recognition speed

Abstract

We propose a technique for regulating LVCSR decoding speed based on a proportional-integral-derivative (PID) model that is widely used in automatic control theory. Our experiments show that such a controller can maintain a given decoding speed level despite computer performance fluctuations, difficult acoustic conditions, or speech material that is out of the scope of the language model, without notable deterioration in overall recognition quality.