A Taxonomy of Specific Problem Classes in Text-to-Speech Synthesis: Comparing Commercial and Open Source Performance

A Taxonomy of Specific Problem Classes in Text-to-Speech Synthesis: Comparing Commercial and Open Source Performance Felix Burkhardt and Uwe D. Reichel Telekom Innovation Laboratories, Research Institute for Linguistics, Hungarian Academy of Sciences Winterfeldstr. 21, 10785 Berlin, Germany, 1068 Budapest VI., Benczur u. 33., Hungary Felix.Burkhardt@telekom.de, uwe.reichel@nytud.mta.hu Abstract Current state-of-the-art speech synthesizers for domain-independent systems still struggle with the challenge of generating understand- able and natural-sounding speech. This is mainly because the pronunciation of words of foreign origin, inflections and compound words often cannot be handled by rules. Furthermore there are too many of these for inclusion in exception dictionaries. We describe an approach to evaluating text-to-speech synthesizers with a subjective listening experiment. The focus is to differentiate between known problem classes for speech synthesizers. The target language is German but we believe that many of the described phenomena are not language specific. We distinguish the following problem categories: Normalization, Foreign linguistics, Natural writing, Language specific and General. Each of them is divided into five to three problem classes. Word lists for each of the above mentioned categories were compiled and synthesized by both a commercial and an open source synthesizer, both being based on the non-uniform unit-selection approach. The synthesized speech was evaluated by human judges using the Speechalyzer toolkit and the results are discussed. It shows that, as expected, the commercial synthesizer performs much better than the open-source one, and especially words of foreign origin were pronounced badly by both systems. Keywords: text-to-speech, evaluation, problem-classes 1. Introduction In this study we are interested in two questions: • To what degree do known “problem classes” for speech synthesizers affect the quality of pronuncia- tion? • What is the difference with respect to the quality of pronunciation between a commercially developed syn- thesizer and an open source development? Current state-of-the-art text to speech synthesizers for domain-independent systems that are based on the non- uniform unit-selection approach still struggle with the chal- lenge of generating understandable and natural-sounding speech. Nonuniform unit selection is the commercially most suc- cessful approach to speech synthesis. It works basically by concatenating best-fitting chunks of speech from large databases, thereby minimizing a double cost function: best fit to neighbor unit and best fit to target prosody. Because signal manipulation is reduced as much as possible, the re- sulting speech sounds most natural (similar to the original speaker) as long as the utterance to synthesize is close to the original domain of the database. Many errors occur because the pronunciation of inflections, compound words and words of foreign origin as well as so-called “non-standard words” (Sproat et al., 2001) often cannot be handled by rules. However, there are too many of these for inclusion in exception dictionaries. Further- more, even if the correct pronunciation would be known to the synthesizer, the necessary syllable combinations are of- ten not present in the acoustic database, leading to audible discontinuities in the resulting output, especially with syn- thesizers based on non-uniform unit-selection. Although the occurrence of each of these hard-to- pronounce …

Content truncated to 500 words