Effects of stress and pronunciation modelling errors on a Turkish HMM based TTS system

The HMM-based text-to-speech (HTS) approach has been increasingly getting more attention from researchers. One of the advantages is the lack of spurious errors that are observed in the unit selection scheme. Here, performance of the first academic Turkish HTS system is reported. Although Turkish is usually read the way it is written, it is not always a one-to-one mapping. Moreover, although Turkish stress markers typically follow rules, it is not always easy to use those rules for accurate stress marking algorithms. Thus, besides the quality of our baseline system, the sensitivity of the system to pronunciation and stress marker errors are also investigated and reported. We have found that using a more complex pronunciation model and stress make the most impact in the sound durations which improve the quality. However, the improvements with the complex pronunciation model and stress do not add up.