Hiroyuki SEGI

To develop emotional speech synthesis technology for sound broadcasting services, listening test judging emotion of speech data was conducted. The results show that about 300 data can be available to formulate the rule of controlling emotional characteristics of speech by extracting the data with more than 70% answered correctly.

Spoken language plays an important role to transmit broadcasting information to many viewers. Speech signal processing is expected to assist elderly people who are hard to understand rapid speech and/or to develop any tools for improving announcer's skill and language learning effect. We have researched "Speech Rate Conversion" and "Voice Conversion". The speech rate conversion technology therefore has built into TV and radio receivers and has applied to new internet radio news service capable of selecting its speech rates and a language learning application for smartphones. Furthermore, analysis and conversion technologies for intonation or accent which are partial functions of voice conversion have applied to a language learning TV program and voice training tools with personal computer. This paper introduces their technical features and the ways how they were implemented.

We have been conducting research on a high-quality speech synthesis system for automatic audio broadcasting. We propose a method for generating manuscripts for speech database to synthesize definite form sentences.

We are developing a tool that can synthesize concatenate word speech and correct its degradation in order to make it broadcasting quality. In this study, several correction functions in were introduced into this tool. It is available to investigate better correction procedure to generate high quality synthesized speech.

It could be useful to apply other speaker's speech data to a concatenative speech synthesis system for complementing lack of original speaker's speech database. As a basic study to investigate ability of phoneme replacement, vowels or long vowels of /CVC/ triphone environment in a continuous speech were replaced by other 24 speakers' ones and perceptual naturalness was evaluated. As a result, it tends to show higher naturalness at 1) specified 3 speakers, 2) length of replaced phoneme is short or 3) F0 of replaced phoneme is lower. Furthermore correlation analysis between evaluated value and spectral feature suggests that 1) there is higher correlation at specified 3 frequency bands and 2) spectral peak or dip is related with evaluated value.

We have been conducting research on a high-quality speech synthesis system for automatic audio broadcasting. We propose voice synthesizer to read out a news flash for visually impaired.

This paper describes our study to implement a service that provides visually impaired people with a read-out presentation of a superimposed news flash text. Data broadcasting is used for automatic start-up of speech that is generated by a speech synthesizer. Four possible solutions are considered and each of them is tested using a trial system.

This paper describes a perceptual experiment on naturalness of replacing a phoneme segment to other speaker's one in a speech synthesis and the results of word speech synthesis by the multi speaker database.

A new pause duration setting method for synthesis by compilation of recorded speech is proposed and its effect is confirmed by subjective evaluation test.

We have been conducting research on a high-quality speech synthesis system for automatic audio broadcasting. We propose stock prices voice synthesizer with numerical speech synthesis method and speech rate conversion.

It is useful to combine multi-speaker's speech database for concatenative speech synthesis system. This paper describes a perceptual study on naturalness and personality by exchanging a phoneme segment to other speaker's one in a word speech synthesis. The experimental result shows some phoneme sequences and speakers are available for the exchangement.

A concatenative speech synthesis method using words and clustered triphones as search units has been proposed. Only 20% words in the sentences, however, satisfy the phoneme context as well as enough quantity in the speech database although most of the words in the sentences can be found in the speech database. So in this paper, we try to use context dependent phoneme sequences not bound by words and to concatenate the speech segments satisfying phoneme context. We produce synthesized speech by proposed method using four different size speech databases. The subjective evaluation on their naturalness shows that 1) mean opinion score is 3.6, 2) enlargement of speech database more than 43.2 hours doesn't contribute the improvement in naturalness 3) the processing time of speech synthesis shows a little increase with enlargement of speech database.

Conventional speech databases of the concatenative speech synthesis systems are prepared by using the sentences read with appropriate speed and reading style for speech synthesis. If we can use broadcast programs in our archives as a speech database, various sentences can be synthesized naturally. So in this paper, we try to use Japanese broadcast news programs as a speech database and use words and triphones as search units in order to search the huge speech database in shorter time. We got 71% preference score in the paired comparison test and 2.6 mean opinion score in five-point MOS test.

A new method for estimating dynamic features used in speech recognition is proposed. The proposed method involves: 1)increasing time resolution of static features used for dynamic feature estimation, 2)relaxing the restriction that estimation span must be an integer multiple of frame shift, and making the estimation span arbitrary. Evaluating the proposed method with a broadcast news testset, word accuracy improved by selecting appropriate estimation span. Error reduction rate in word accuracy was up to 24% relatively in noisy news speech. The overall error reduction rate was 6.5% relatively.

A new method for estimating dynamic features used in speech recognition is proposed. The proposed method involves:1)increasing time resolution of static features used for dynamic feature estimation, 2)relaxing the restriction that estimation span must be an integer multiple of frame shift, and making the estimation span arbitrary. Evaluating the proposed method with a broadcast news testset, word accuracy improved by selecting appropriate estimation span. Error reduction rate in word accuracy was up to 24% relatively in noisy news speech. The overall error reduction rate was 6.5% relatively.

A new method for estimating dynamic features used in speech recognition is proposed. The proposed method involves : 1)increasing time resolution of static features used for dynamic feature estimation, 2)relaxing the restriction that estimation span must be an integer multiple of frame shift, and making the estimation span arbitrary. Evaluating the proposed method with a broadcast news testset, word accuracy improved by selecting appropriate estimation span. Error reduction rate in word accuracy was up to 24% relatively in noisy news speech. The overall error reductionrate was 6.5% re1atively.

We are studying speech recognition for simultaneous subtitling of news programs. One of our current studies is recognition of reporter's speech in the afflicted area, from overseas and so on. Because these speeches are uttered in various acoustic conditions, we need technologies to handle such conditions. In this paper, a filter-bank subtraction technique is proposed for robust speech recognition under various acoustic conditions. It calculates, in each pass band of the filter-bank, the minimum value of the filter output in a certain interval of time, and then estimates a noise spectrum over all bands as a composition of those minimum values. The experiments show that the new method yielded better recognition performance than the conventional spectral subtraction techniques.

This paper proposes a method of constructing acoustic models for automatic news subtitling. In this method, two-stage clustering is used for an adaptation-data selection from a large-scale database. The first stage generates GMMs from a small amount of data and second atage clusters a large-scale database by the GMMs. A broadcast news transcription experiment showed that the proposed method reduced in toth a word error rate and processing time. The best result in the experiment showed 33% WER reduction and 25% decoding time reduction compared with the baseline method, which had no clustering. Furthermore, we found that it takes just 0.4 seconds from the beginning of an input utterance to select an HMM correctly. The result showed 18% processing time reduction in total compared with the baseline method.

In this paper, the effects of the various acoustic conditions on the speech recognition are reported. A broadcast transcription system for simultaneous subtitling of news program is being developed. Speech recognition for studio announcers' speech, especially for announcers' read speech, has been investigated so far. To realize simultaneous subtitling for the whole news program, speech in various acoustic conditions should be recognized. For example, speech spoken by people other than announcers or speech with background sound such as noise and music become the recognition object. Speech in the whole news program was recognized to investigate the effort of the acoustic condition to the recognition accuracy.