Recent developments in speech synthesis have produced systems capable of producing speech which closely resembles natural speech, and researchers now strive to create models that more accurately mimic human voices. One such development is the incorporation of multiple linguistic styles in various languages and accents. Speech synthesis based on Hidden Markov Models (HMM) is of great interest to researchers, due to its ability to produce sophisticated features with a small footprint. Despite some progress, its quality has not yet reached the level of the current predominant unit-selection approaches, which select and concatenate recordings of real speech, and work has been conducted to try to improve HMM-based systems. In this paper, we present an application of long short-term memory (LSTM) deep neural networks as a postfiltering step in HMM-based speech synthesis. Our motivation stems from a similar desire to obtain characteristics which are closer to those of natural speech. The paper analyzes four types of postfilters obtained using five voices, which range from a single postfilter to enhance all the parameters, to a multi-stream proposal which separately enhances groups of parameters. The different proposals are evaluated using three objective measures and are statistically compared to determine any significance between them. The results described in the paper indicate that HMM-based voices can be enhanced using this approach, specially for the multi-stream postfilters on the considered objective measures.
Measuring the perceptual effects of modelling assumptions in speech synthesis using stimuli constructed from repeated natural speech
