This service was developed to provide a natural speech fillers feature to the DO voicebot. It works by modifying the GDO's reponse by inserting natural speech fillers such as "uh", "um" and pauses. The aim of this service is to give the chatbot a more natural tone of voice. The number of fillers inserted can be controlled with the help of a level_of_naturalisation variable.
This service is not fully implemented. While filler generation and insertion works, the text-to-speech service is not adapted to process the reponse from the speech fillers service. As of now the speech filler service returns as output the modified text with included vocal tags which can be handled by Cereproc.
To perform fillers insertion we utilise data taken from the Movie Dialog Corpus which can be found 'here'.
We pre-processed the dataset taken from the Movie Dialog Corpus, by substituting all variants of ‘um’ with it and we did the same for ‘uh’. We artificially increased the amount of filler pauses in the data by replacing long dashes and ellipses with a filler pause as well.
The method we used for the insertion of natural speech fillers into a sentence is based on the bigram method proposed by Sharma, Shah and Joshi in the paper "Naturalization of Text by the Insertion of Pauses and Filler Words" which can be read 'here'.