Tham khảo tài liệu 'recent advances in signal processing 2011 part 11', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | Estimation of the instantaneous harmonic parameters of speech 337 b c Fig. 8. Harmonic parameters estimation a source signal b estimated deterministic part c estimated stochastic part An example of harmonic analysis is presented in Figure 8 a . The source signal is a phrase uttered by a male speaker Fs 8kHz . The deterministic part of the signal Figure 8 b was synthesized using estimated harmonic parameters and subtracted from the source in order to get the stochastic part Figure 9 c . The spectrograms show that all steady harmonics of the source are modelled by sinusoidal representation when the residual part contains transient and noise components. Harmonic analysis in TTS systems This subsection presents an experimental application of sinusoidal modelling with proposed analysis techniques to a TTS system. Despite the fact that many different techniques have been proposed segment concatenation is still the major approach to speech synthesis. The speech segments allophones are assembled into synthetic speech and this process involves time-scale and pitch-scale modifications in order to produce natural-like sounds. The concatenation can be carried out either in time or frequency domain. Most time domain techniques are similar to the Pitch-Synchronous Overlap and Add method PSOLA Moulines and Charpentier 1990 . The speech waveform is separated into short-time signals by the analysis pitch-marks that are defined by the source pitch contour and then processed and joined by the synthesis pitch-marks that are defined by the target pitch contour . The process requires accurate pitch estimation of the source waveform. Placing 338 Recent Advances in Signal Processing analysis pitch-marks is an important stage that significantly affects synthesis quality. Frequency domain parametric techniques deal with frequency representations of the segments instead of their waveforms what requires prior transformation of the acoustic database to frequency domain. Harmonic modelling
