Robust speech-based happiness recognition

摘要

This paper presents a robust happiness recognition system. The system consists of a happiness recognition module and a noise suppression module. In the happiness recognition module, we present an emotion feature set comprising Mel-frequency cepstral coefficients (MFCCs), the subband powers, spectral centroid, spectral spread, spectral flatness, RSS, pitch and energy. The proposed feature set is fed into a probability product support vector machine for happiness recognition. In real world applications, the speech received are often exposed to noise, thus prone to reducing the recognition rate. We propose a noise suppression method using subspace based method. A gain function estimation method is used for time domain constrained (TDC) based subspace speech enhancement. The optimal Lagrange multiplier of the gain function will be estimated in accordance with signal to noise ratio (SNR) of the noisy speech. The proposed happiness recognition system has been tested using a large number of noisy speech utterances with a 34% equal error rate.