Concatenative speech synthesis: a Framework for Reducing Perceived Distortion when using the TD-PSOLA Algorithm.

Abstract

This thesis presents the design and evaluation of an approach to concatenative speech synthesis using the Titne-Domain Pitch-Synchronous OverLap-Add (I'D-PSOLA) signal processing algorithm. Concatenative synthesis systems make use of pre-recorded speech segments stored in a speech corpus. At synthesis time, the `best' segments available to synthesise the new utterances are chosen from the corpus using a process known as unit selection. During the synthesis process, the pitch and duration of these segments may be modified to generate the desired prosody. The TD-PSOLA algorithm provides an efficient and essentially successful solution to perform these modifications, although some perceptible distortion, in the form of `buzzyness', may be introduced into the speech signal. Despite the popularity of the TD-PSOLA algorithm, little formal research has been undertaken to address this recognised problem of distortion. The approach in the thesis has been developed towards reducing the perceived distortion that is introduced when TD-PSOLA is applied to speech. To investigate the occurrence of this distortion, a psychoacoustic evaluation of the effect of pitch modification using the TD-PSOLA algorithm is presented. Subjective experiments in the form of a set of listening tests were undertaken using word-level stimuli that had been manipulated using TD-PSOLA. The data collected from these experiments were analysed for patterns of co- occurrence or correlations to investigate where this distortion may occur. From this, parameters were identified which may have contributed to increased distortion. These parameters were concerned with the relationship between the spectral content of individual phonemes, the extent of pitch manipulation, and aspects of the original recordings. Based on these results, a framework was designed for use in conjunction with TD-PSOLA to minimise the possible causes of distortion. The framework consisted of a novel speech corpus design, a signal processing distortion measure, and a selection process for especially problematic phonemes. Rather than phonetically balanced, the corpus is balanced to the needs of the signal processing algorithm, containing more of the adversely affected phonemes. The aim is to reduce the potential extent of pitch modification of such segments, and hence produce synthetic speech with less perceptible distortion. The signal processingdistortion measure was developed to allow the prediction of perceptible distortion in pitch-modified speech. Different weightings were estimated for individual phonemes,trained using the experimental data collected during the listening tests.The potential benefit of such a measure for existing unit selection processes in a corpus-based system using TD-PSOLA is illustrated. Finally, the special-case selection process was developed for highly problematic voiced fricative phonemes to minimise the occurrence of perceived distortion in these segments. The success of the framework, in terms of generating synthetic speech with reduced distortion, was evaluated. A listening test showed that the TD-PSOLA balanced speech corpus may be capable of generating pitch-modified synthetic sentences with significantly less distortion than those generated using a typical phonetically balanced corpus. The voiced fricative selection process was also shown to produce pitch-modified versions of these phonemes with less perceived distortion than a standard selection process. The listening test then indicated that the signal processing distortion measure was able to predict the resulting amount of distortion at the sentence-level after the application of TD-PSOLA, suggesting that it may be beneficial to include such a measure in existing unit selection processes. The framework was found to be capable of producing speech with reduced perceptible distortion in certain situations, although the effects seen at the sentence-level were less than those seen in the previous investigative experiments that made use of word-level stimuli. This suggeststhat the effect of the TD-PSOLA algorithm cannot always be easily anticipated due to the highly dynamic nature of speech, and that the reduction of perceptible distortion in TD-PSOLA-modified speech remains a challenge to the speech community.