Cross-Gender and Age Speech Conversion Using Hidden Markov Model Based on Cepstral Coefficients Conversion

Animation movies often use children’s characters and they need children aged 5–10 to do a dubbing. For cost efficiency, a speech conversion can be done to support dubbing a children’s speech. To deal with it, in this research we propose the method to converting an adult’s speech to children’s speech. The contribution of this study is to design a signal processing algorithm to perform the conversion. In this study we propose a conversion method using the Hidden Markov Model (HMM) based on Cepstral Coefficients Conversion. The input is the speech of source speakers and the target speakers that using similar sentences. Features extraction, which is used is by extracted pitch (f₀) and cepstral in conversion process, and the modeling method is HMM. System output is converted speech signals that has similar characteristics with target speech signal. From the testing results, the most optimal HMM parameter is using 4-state. The highest increase of cepstral Root Mean Square Error (RMSE) before conversion and after conversion is equal to 32.35% and an average 25.83% which obtained from 400 samples. Mean Opinion Score (MOS) on a scale from 1 (converted speech is very dissimilar with the target speech) to 5 (converted speech is very similar with the target speech). It resulted an average value of 2.505 in terms of similarities and has an average value of 2.805 in terms of quality which obtained from 30 respondents. The proposed method is expected to be used in the animation film industry in order to simplify and make efficient the dubbing process.