Taro Togawa

Source separation based on transfer function between microphones and its dispersion

In this paper, we propose a technique for realizing the source separation required when analyzing face-to-face conversations between multiple speakers. In the conventional method based on binary masking, it is assumed that the input signal from the microphone nearest to the speaker is stronger than that from the other microphones. However, due to reflection, this assumption may not necessarily be true. In our proposed method, we estimate the transfer function between microphones based on the spectrum difference in the segment when each person is speaking, and we control the amount of suppression required for each input signal based on the estimated transfer function between the microphones and its dispersion. The experimental results for a simulated environment of face-to-face conversations shows that the proposed method suppresses the interference signal by 27.0 dB and distorts the target signal by only 1.5 dB. These results demonstrate that the proposed method is effective in environments in which reflection occurs due to obstacles.

Development of speech technologies to support hearing through mobile terminal users

Mobile terminals have become the most familiar communication tool we use, and various types of people have come to use mobile terminals in various environments. Accordingly, situations in which we talk over the telephone in noisy environments or with someone who speaks fast have increased. However, it is sometimes difficult to hear a persons voice in these cases. To make the voice received through mobile terminals easy to hear, authors have developed two technologies. One is a voice enhancement technology that emphasizes a callers voice according to the noise surrounding the recipient, and the other is a speech rate conversion technology that slows speech while maintaining voice quality. In this paper, we explain the trends and the features of these technologies and discuss ways to implement their algorithms on mobile terminals.

Pitch conversion method and device for converting a pitch of an input signal into a desired pitch

In a pitch conversion method and device which can reduce data throughput while suppressing a degradation of sound quality due to a pitch conversion as much as possible, an input signal pitch pattern per predetermined processing unit and a target pitch pattern are inputted, and a degradation degree indicating how a waveform of the input signal degrades upon pitch conversion from the input signal pitch pattern to the target pitch pattern is calculated. Alternatively, a degradation degree corresponding to a voice state and a phonemic type of the input signal is extracted from a database in which all of combinations of voice states and phonemic types estimated are associated with the degradation degrees to be recorded. Then, a pitch converter which performs a pitch conversion with small data throughput and a pitch converter which performs a pitch conversion with large data throughput are switched over depending on the degradation degree.