JPS58223198A - Syllable inputting system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention uses a syllable input method (:) that recognizes a single syllable (for example, two uttered sounds) and outputs a kana character string, and a syllable input method (: that detects a speech interval from an input back door). This is related to two things.

<Conventional technique> The conventional syllable input method (:
The input voice is calculated for each frame (Niseido power a is calculated, and compared with a threshold value predetermined from the unmanned power.

As a result of the comparison, if frames of a predetermined length (n, frames) or more are continuous, the frame in which the voice power exceeds the threshold is set as the start point C of the voice section. After detecting the start of a voice section, if frames with voice power below the threshold continue for more than a predetermined length of frames (NLC frames), a silent section is detected and the voice power is below the threshold: The frame in which this occurs is set as the end d of the audio section.

In the conventional method described above, after the utterance of a syllable, the interval (nB
The problem was that the end of a syllable could not be detected unless the next syllable was uttered after a gap of at least one frame. Therefore, it is not possible to pronounce syllables one after another, and the user is forced to pause between syllables, which is a heavy burden, and the pause between the previous clauses is long, so the speech production speed is faster than normal speech production. It was too late and it was difficult to speak.

<Summary of the Invention> This invention solves these drawbacks by: 1. By detecting the end of a voice section using not only the information on the asexual section but also the information on the valley of voice power. , intersyllabic (
2. It is now possible to detect voice sections without long pauses,
It is possible to input syllables at high speed, and although each syllable is uttered separately, there is no need to insert pauses, and it is possible to input speech at about the same speed as normal speech.

<Embodiment> FIG. 2 is a block diagram showing the configuration of an embodiment of the present invention. The audio input signal from the microphone 1 passes through a bandpass filter 2 and is band-limited, and then converted into a digital signal 12 by an AD converter 3 (2).

Next, the audio power of this digital signal (M) is calculated by the audio power calculation unit 4C2 with a predetermined frame length and frame period, and is output to the audio power output terminal 5.

The silent section detection unit 6 compares the threshold value set by the unmanned power C2 with the audio power input from the audio power output terminal 5 for each frame (:
When the audio power is greater than the threshold value, "1" is output, and when the audio power is less than the threshold value, "θ" is output from the silent section output terminal 7.

A power dip (trough in audio power) detection section 8 detects a trough in the audio power input through the audio power output terminal 5. As a power dip detection method, it can be considered that it is included in the package, but an example is the second-order differential coefficient when approximating the quadratic curve of the audio power time series. Compare each frame with an appropriate threshold value set in advance (-, and if the differential coefficient is larger than the threshold value, C- is “
l”, if smaller than the threshold, power 10”;
Tsubu output terminal 9 (:, also outputs the value of the differential coefficient to the differential coefficient output terminal 10 (-).

The back door section determining section 11 connects the silent section output terminal 7 and the power
The start and end of the voice section are determined from information input through the dip output terminal 9 and the differential coefficient output terminal 10. As a specific example of the method, there are two methods described below. Namely, (1) the value of the silent section output terminal 7 is from 101 to 81 m1 (
After the duplexing, if frames of ``'1'' continue for nB frames or more, the frame converted from 10'' to ``11'' is taken as the start of the audio.

■) After detecting the start of the sound, the value of the silent section output terminal 7 becomes 1.
1# to @0' 4m change, 0# frame is 06
If more than one frame continues (:, then ``1'' to ``
The frame that changes to θ''' is the end of the audio.

(3) After detecting the start of the audio, the period where the value of the silent section output terminal 7 is 10" is one frame or more and less than nE frames, and the value of the silent section output terminal 7 is @0"(; power When there is a frame in which the value of the dip output terminal 9 is ``1#'', the frame in which the value of the silent section output terminal 7 changes from 1# to "O" is set as the end of the audio.

(4) After detecting the start of the voice, when the value of the silent section output terminal 7 is "1'" and the value of the power dip output 91M child 9 is "1", the value of the power dip output terminal 9 is 11m.
During the interval, the frame with the maximum differential coefficient value input from the differential coefficient output terminal 10 is set as the end of the speech and the beginning of the next syllable.

The purpose of this invention is not only to have a silent section, but also to make the valley of the audio power the end of the sound. Instead, a similar principle may be used.Since each syllable (: voice is input), even if adjacent syllables are close to each other, the second-order differential coefficient of the speech power between syllables will increase.

The audio section detected using the above method is sound 1106 Oshibatake 1
2, the kana character string is output from the NM recognition button output terminal 13. For example, Nakatsu,
Hamada, Hoben, Manhama, "Study of the Japanese Journal Monosyllabic Speech Introduction Method," 1988 J56 Anti-electronics Society Information Systems Division National Conference Preliminary Collection 1-117 (October 1982) .

FIG. 3 is an example of the above-described method (2). chi j issue e
is required. In addition, the quadratic differential coefficient f when the audio power obtained in the power dip detection 6118 is approximated by a quadratic curve is outputted from the differential coefficient output terminal 1o (== and the comparison with the threshold value g≦2 shows that the power dip is The output signal is determined. From this information, the start point C1 and end point d of the voice section are determined based on the above-mentioned information.

As explained above, this invention has the advantage that it is not necessary to provide a silent section for each dorsal intersegmental, since the present invention uses not only silent section detection but also the trough information of the voice power to detect the end of speech. Therefore, each phrase 11'O in a word or phrase is pseudo-continuous (the user can input it in any rapid response, such as by saying it twice).
:A fast-paced input method can be realized.

[Brief explanation of the drawing]

Fig. 1 is a waveform diagram showing an example of speech interval detection in the conventional 100-section input method (:), Fig. 2 is a block diagram showing case 1 戊 according to an embodiment of the present invention, and Fig. 3 It is a waveform diagram showing the E/IJ of voice section detection in the Kubikiri Kaka type (-). Microphone, 2: band pass filter, 3: AD converter, 4: voice power calculation unit, 5: voice power output terminal , 6:
Silent section detection section, 7: Silent section output terminal, 8: Power/dip detection section, 9: Power/dip output terminal, lO: Differential coefficient output terminal, 11: Voice section determining section, 12: Bamboo knot recognition section, 13: Recognition result output terminal. Patent applicant: Japan Aribo Telephone Corporation

Claims (1)

[Claims] (1) For each sound h0 1: A syllable input method that recognizes each syllable of the uttered voice (: A method for calculating the power of the input voice every frame of a predetermined length 4: A means for calculating the power of the input voice and its expansion power means for detecting a silent interval from 62 by comparing with a predetermined threshold; means for determining the end of a syllable at the valley of the voice power detected from the voice power calculated above; Detected syllable end (
- A syllable input method characterized by comprising means for detecting speech intervals.