著者

重永 実 有泉 均

出版者

一般社団法人 日本音響学会

雑誌

日本音響学会誌 (ISSN:03694232)

巻号頁・発行日

vol.21, no.5, pp.263-271, 1965-09-30 (Released:2017-06-02)

---

A method of recognizing /m/ and /n/ of monosyllables and words in real time is reported. The method consists of three parts; segmentation of nasal consonants, recognition of the following vowel, and discrimination between /m/ and /n/ according to the result of vowel recognition. In order to extract the nasal part, /e, a, o, u, w/ are excluded from nasals by comparing the output of 300 c/s LPF with that of 500〜1600 c/s BPF, and /i, j/ are excluded by comparing 500〜1600 c/s with 2800〜5000 c/s. Voiceless stops are easily omitted by comparing the output of 300 c/s LPF with that of 700 c/s HPF, and this circuit is also used for excluding vowels. In order to exclude the beginning and end part of /u/ more certainly, the output wave form of 400〜1000 c/s BPF is used. The parts in which the envelope of that wave varies rapidly are excluded from the nasal part. For excluding voiced stops, fricatives, and flappeds, fundamental frquency components are extracted from filtering output of 400〜1000 c/s BPF and the parts in which the fundamental frequency components exist continuously are considered to be likely nasal. The low intensity parts of original speech waves are considered to be non-nasal parts. The segment which satisfies above six conditions is decided to be nasal consonant after leaving out 12 msec of the continuum. According to this method the initial part of the nasal consonant is often missed but the boundary between the nasal consonant and the following vowel is pointed out exactly. Concerning about the discrimination between /m/ and /n/, the components of two specific frequencies are compared just before the boundary between the nasal consonant and the following vowel for distinguishing /mi/ and /me/ from /ni/ and /ne/ respectively. For the nasals followed by /a, o, u/, F_2 loci are used directly. Though /me/ and /ne/ in words are not discriminated satisfactorily by the method due to the individual variations, the others are recognized correctly over 80% for 180 samples of three male speakers.

著者

神谷 賢 雨宮 沙織 有泉 均

出版者

一般社団法人電子情報通信学会

雑誌

電子情報通信学会技術研究報告. SP, 音声 (ISSN:09135685)

巻号頁・発行日

vol.101, no.86, pp.7-13, 2001-05-17

被引用文献数

1

---

筆者らは声道模擬型の規則音声合成器の研究をしている. 今回人間の放射音声/a/から求めた逆フィルタ波形を1波形に切り出さないで連続したまま音声合成器へ入力した. すると, その人の個人性を保存した肉声レベルの/aiueo/を合成できた. このため, この方法を使えば, 数式でのモデル化が困難な声質も実現できる可能性がある. また, 筆者らの合成方法はターミナルアナログ型であるため, 人間の発声機構の理解に役立つと思われる. 今後は様々な性質を実現する.