Pitch Marking Based on an Adaptable Filter and a Peak-Valley Estimation Method

1 Pitch Marking Based on an Adaptable Filter and a Peak-Valley Estimation Method Jau-Hung Chen and Yung-An Kao Advanced Technology Center, Computer and Communication Research Laboratories, Industrial Technology Research Institute, Chutung 310, Taiwan Email: chenjh@itri.org.tw, kya@itri.org.tw Abstract In a text-to-speech (TTS) conversion system based on the time-domain pitch-synchronous overlap-add (TD-PSOLA) method, accurate estimation of pitch periods and pitch marks is necessary for pitch modification to assure an optimal quality of the synthetic speech. In general, there are two major issues on pitch marking: pitch detection and location determination. In this paper, an adaptable filter, which serves as a bandpass filter, is proposed for pitch detection to transform the voiced speech into a sine-like wave. Based on the sine-like wave, a peak-valley decision method is investigated to determine the appropriate part (positive part and negative part) of the voiced speech for pitch mark estimation. At each pitch period, two possible peaks/valleys are searched and the dynamic programming is performed to obtain the pitch marks. Experimental results indicate that our proposed method performed very well if correct pitch information is estimated. 1. Introduction In past years, the approach of concatenative synthesis has been adopted by many text-to-speech (TTS) systems [1]–[6]. The concatenative synthesis uses real recorded speech segments as the synthesis units and concatenates them together during synthesis. Also, the time-domain pitch-synchronous overlap-add (TD-PSOLA) [6] method has been employed to perform prosody modification. This method modifies the prosodic features of the synthesis unit according to the target prosodic information. Generally, the prosodic information of the speech includes pitch (the fundamental frequency), intensity, and duration, etc. For a synthesis scheme based on TD-PSOLA method, it is necessary to obtain a pitch mark for each pitch period in order to assure an optimal quality of the synthetic speech. The pitch mark is a reference point for the overlap of the speech signals. 2 It is useful to have a speech synthesizer with various voices for speech synthesis. Sometimes it is also important for a service-providing company to have a synthesizer with the voice of its own employee or the speaker of its favorite. For conventional TTS systems, however, it is a professional but tedious job to create a new voice. Recently, corpus-based TTS systems have been appreciated which use a large amount of speech segments. Some approaches selected the speech segments as the candidates of synthesis units. Establishing the synthesis units includes speech segmentation, pitch estimation, pitch marking, and so on. However, pitch marking is very labor-intensive among them if there involved no automatic mechanism. In general, there are two major issues on pitch marking: pitch detection and location determination. Compared to pitch detection [7]-[14], few papers have been presented for pitch marking [15][16], which is also a difficult problem because of the great variability of the speech signals. Moulines et al. [15] proposed a pitch-marking algorithm based on the detection of abrupt changes at glottal closure instants. At each period, they assumed that the speech waveform could be represented by the concatenation of …

Content truncated to 500 words