Influence of Emotional Speech on Continuous Speech Recognition

Emotions are an important part of human communication, but they can present harsh conditions for an automatic continuous speech recognition system. This paper presents an analysis of to which level the emotional speech degrades speech recognition accuracy, when dealing with a highly inflected Slovenian language. Namely, the language characteristics are those that also influence the speech recognition performance, and inflection is one of the most challenging ones. Moreover, Slovenian belongs to the group of under-resourced languages, like other Slavic languages. The speech recognition system was developed with the Slovenian BNSI Broadcast News speech database. The Interface speech database was used for the experiments with the emotional speech. The analysis was carried out with HMM and DNN acoustic models, combined with a 3-gram statistical language model. The results show that emotional speech degrades speech recognition accuracy in the range between 5% and 7% absolutely.