Improving emotion detection with speech by enhanced approach

Emotion detection currently is found to be an important and interesting part of speech analysis. The analysis can be done by selection of an effective parameter or by combination of a number of parameters to gain higher accuracy level. Definitely selection of a number of parameters together will provide a reliable solution for getting higher level of accuracy than that of for the single parameter. Energy, MFCCs, pitch values, timbre, vocal tract frequencies are found to be effective parameters with which detection accuracy can be improved. The work presented here is by use of database with one of the Indian language Marathi of 1200 speech files recorded by male and female professional actors for six different emotions happy, angry, neutral, sad, fear and surprised. It is noted that results with the language are proportional with results of other languages indicating that language is an independent parameter for emotion detection. Similarly, by addition of an effective classifier like neural network can further yield the recognition accuracy nearly to 100%. The work attempts to interpret the fact that combining the results of each parameter have improved the detection accuracy. In near future, it may also lead to design a system capable to work by understanding human emotions.