Search

Differences in speaker characteristics, recording conditions, and signal processing algorithms affect output quality in voice conversion systems. This study focuses on formulating robust techniques for a codebook mapping based voice conversion algorithm. Three different methods are used to improve voice conversion performance: confidence measures, pre-emphasis, and spectral equalization. Analysis is performed for each method and the implementation details are discussed. The first method employs confidence measures in the training stage to eliminate problematic pairs of source and target speech units that might result from possible misalignments, speaking style differences or pronunciation variations. Four confidence measures are developed based on the spectral distance, fundamental frequency (f0) distance, energy distance, and duration distance between the source and target speech units. The second method focuses on the importance of pre-emphasis in line-spectral frequency (LSF) based vocal tract modeling and transformation. The last method, spectral equalization, is aimed at reducing the differences in the source and target long-term spectra when the source and target recording conditions are significantly different. The voice conversion algorithm that employs the proposed techniques is compared with the baseline voice conversion algorithm with objective tests as well as three subjective listening tests. First, similarity to the target voice is evaluated in a subjective listening test and it is shown that the proposed algorithm improves similarity to the target voice by 23.0%. An ABX test is performed and the proposed algorithm is preferred over the baseline algorithm by 76.4%. In the third test, the two algorithms are compared in terms of the subjective quality of the voice conversion output. The proposed algorithm improves the subjective output quality by 46.8% in terms of mean opinion score (MOS).

The quality of text-to-speech systems can be effectively assessed only on the basis of reliable and valid listening tests to assess overall system performance. A mean opinion scale (MOS) has been the recommended measure of synthesized speech quality [ITU-T Recommendation P.85, 1994. Telephone transmission quality subjective opinion tests. A method for subjective performance assessment of the quality of speech voice output devices]. We assessed this MOS scale and developed and tested a modified measure of speech quality. This modified measure has new items specific to text-to-speech systems. Our research was motivated by the lack of clear evidence of the conceptual content of as well as the psychometric properties of the MOS scale. We present conceptual arguments and empirical evidence for the reliability and validity of a modified scale. Moreover, we employ state of the art psychometric techniques such as confirmatory factor analysis to provide strong tests of psychometric properties. This modified scale is better suited to appraise synthesis systems since it includes items that are specific to the artifacts found in synthesized speech. We believe that the speech synthesis research communities will find this modified scale a better fit for listening tests to assess synthesized speech.

This paper describes perceptual methods for diagnosing problems in text-to-speech systems. Special attention is paid to two issues. First, coverage of the domain of a text-to-speech system. Since this domain involves an enormous range of contexts, it is criticial for diagnostics, and also for overall evaluation, that test materials cover this range to the fullest extent possible. Automatic text generation algorithms that make extensive use of "greedy" algorithms are described that serve this purpose. Second, speech generated by text-to-speech systems tends to have a great variety of problems . A battery of experimental paradigms is discussed that address different facets of speech quality and intelligibility. Included are: (a) "word pointing" method for detection of problematic concatenative units, (b) "minimal pairs intelligibility test"—an expanded diagnostic rhyme test; (c) automatically scored orthographic name transcription task; (d) mean opinion score paradigm with problem categorization; and (e) paired comparison paradigm with strength-of-choice rating. The methods are applied in a series of experiments on high-end text-to-speech systems.