世木 寛之

Tied-mixture HMMs have been proposed as the acoustic model for large-vocabulary continuous speech recognition and have yielded promising results. They share base-distribution and provide more flexibility in choosing the degree of tying than state-clustered HMMs. However, it is unclear which acoustic models to superior to the other under the same training data. Moreover, LBG algorithm and EM algorithm, which are the usual training methods for HMMs, have not been compared. Therefore in this paper, the recognition performance of the respective HMMs and the respective training methods are compared under the same condition. It was found that the number of parameters and the word error rate for both HMMs are equivalent when the number of codebooks is sufficiently large. It was also found that training method using the LBG algorithm achieves a 90% reduction in training time compared to training method using the EM algorithm, without degradation of recognition accuracy.

The 'Kabushiki Shikyo' program broadcast on NHK Radio 2 reports on the daily closing prices and net changes of about 830 stocks listed on the Tokyo Stock Exchange. Reading out the numerical values without making mistakes within the allotted broadcast time can be extremely difficult for the announcers. We have therefore developed an automatic broadcast system for stock-price bulletins, which uses numerical speech synthesis and automatic speech-rate conversion. Our system has been used in experimental digital terrestrial radio broadcasts since October 2006 and also used in NHK radio 2 since March 2010. This article describes the generation of texts to build the speech waveform database, the mechanism used to synthesize numerical speech via the database, and the evaluation of naturalness for the synthesized speech samples.

Tied-mixture HMMs have been proposed as the acoustic model for large-vocabulary continuous speech recognition and have yielded promising results. They share base-distribution and provide more flexibility in choosing the degree of tying than state-clustered HMMs. However, it is unclear which acoustic models to superior to the other under the same training data. Moreover, LBG algorithm and EM algorithm, which are the usual training methods for HMMs, have not been compared. Therefore in this paper, the recognition performance of the respective HMMs and the respective training methods are compared under the same condition. It was found that the number of parameters and the word error rate for both HMMs are equivalent when the number of codebooks is sufficiently large. It was also found that training method using the LBG algorithm achieves a 90% reduction in training time compared to training method using the EM algorithm, without degradation of recognition accuracy.