Floating to Fixed-Point Translation with Its Application to Speech-Based Emotion Recognition

摘要

Speech-based emotion recognition is one of the latest challenges in speech processing. The algorithms are developed using floating-point arithmetic because of its wide dynamic range and constant relative accuracy. However, they are finally implemented in hand held devices which are required to consume less power, time and have a lower market price. Fixed-point arithmetic with proper determination of integer and fractional bitwidths can help in satisfying these requirements. Therefore we have made an attempt to develop a fixed-point speech-based emotion recognition system using Mel frequency cepstral coefficients (MFCCs) and hidden Markov model (HMM). Accurate range and precision analysis has been carried out to compute optimum integer and fractional word lengths. The speech emotion engine has been evaluated using Berlin emotional speech database, EMO-DB. A speaker independent emotion recognition accuracy of 71.02% and 67.42% for floating-point and fixed-point formats with optimized wordlenghs respectively was achieved. Finite wordlength effect like quantization with range of relative errors and its effect on emotion recognition task has been analyzed.