JPH039400A - Voice recognizer - Google Patents

Abstract

PURPOSE:To improve the recognition rate in both low-noise environment and high-noise environment by indicating the loudness of a voice to a speaker according to the loudness of a background noise, and switching and using two kinds of storage means which are prepared. CONSTITUTION:A noise detecting means 108 measures and decides the background noise to detect how loud the noise is, and indicates normal or loud voice input to a speaker according to the detection result. Then when the voice input of the speaker is recognized, a 1st storage means 105 or 2nd storage means 106 is used selectively. For example, when it is judged that a noise is low, feature parameters stored in the 1st storage means 105 are utilized to recognize an input voice. When the noise is loud, on the other hand, feature parameters stored in the 2nd storage means 106 are utilized to recognize the input voice. Consequently, the excellent recognition rate is obtained in both the low-noise environment and loud-noise environment.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device, and particularly to speech recognition in a high noise environment.

[Prior Art] Usually, speech recognition analyzes the speech produced by a speaker,
The similarity with standard patterns (also called a dictionary) registered in advance is calculated, and the most similar standard pattern is output as a recognized pattern.

Figure 2 is by Furui [Digital Audio Processing] (September 1985).
125) is a block diagram of a speech recognition device when assuming word-by-word speech input, which is disclosed in Tokai Inu q Dezakakai, P. 160-181. In the figure, (201)
is an input terminal, (202) is a voice analysis means, and the input terminal (
20+,') is subjected to necessary processing to convert it into feature parameters that are effective for recognition.

(203) is a memory in which standard patterns of words are stored, and characteristic parameters of words to be recognized are determined by sound analysis means (202). I) Required and stored. (204) is a word recognition means, and the voice to be the subject of the four words is similarly processed 17 by the voice analysis means (202).
The feature parameters derived by
The similarity with all the standard patterns of the stored words is calculated, and the pattern with the highest degree of similarity (similarity) is outputted to the output terminal (205) as a recognition result.

FIG. 3 is a diagram showing the operation of the one-high voice recognition device shown in FIG.
Char 1. & When voice recognition starts (S301)
, microphone (not shown), etc.
1) Input voice (S302). Voice analysis means (2
02) derives the feature parameters after analyzing the voice performance 1.5. Then, the word recognition means (203) uses the characteristic parameters from the speech analysis means (202) and the memory (203).
18. Calculate the similarity with all word standard patterns stored in 18. Then, the pattern with the highest degree of similarity is determined as the recognition result (S3 (13), and output L to the output terminal (205).
(S304), and the process ends (S305).

Usually, a speech analysis stage (202) and a word recognition means (20
4) is realized using a digital signal processor (DSP), etc., so the audio input to the input terminal (201) is amplified by inputting the audio signal from the microphone and then subjected to appropriate filter processing. After 1: A /
This is D-converted digital data. Also,
The recognition result is encoded into an appropriate code and output to the control unit of the controlled device via the output terminal (205). For example, if it is a voice typewriter, it will be output to a circuit that controls the typewriter's ^ and sod, and if it is a voice dial, it will be output to a circuit that generates dial pulses or button tones.

In some cases, echo back to the speaker may also be echoed using a disb (noise, voice synthesizer, etc.) in parallel.

[Problem to be solved by the invention 12] The voice recognition device described in 1 above cannot be used in an environment with large background noise, such as in a car while driving at high speed.
There may be background noise that is equal to or greater than the voice uttered by the speaker (i.e., when the S/pause is negative).
However, there is a problem in that voice analysis and recognition cannot be performed accurately.

In addition, in such a high-noise environment F, 'C1 humans naturally vocalize in a loud voice 1. There is a tendency to try.

Speak out loud1. In this case, the S/N ratio is improved, but on the other hand, the characteristic parameters of the voice are often different from the normal characteristic parameters, and there is still the problem that accurate recognition cannot be performed.

The present invention aims to solve the conventional problem of a decrease in recognition rate in a high noise environment and to provide a speech recognition device that can obtain a good recognition rate in both a low noise environment and a ρ1 noise environment. purpose.

[Means for Solving the Problems] A speech synthesis device according to the present invention includes a first storage means in which characteristic parameters of normal speech are stored, and a second storage means in which characteristic parameters of large speech are stored. , a noise detection means for measuring and determining background noise and detecting its magnitude;
It has a voice synthesizing means for instructing the speaker whether to input the voice in normal μj or in a loud voice according to the output of the miscellaneous δl novel detection means.

Furthermore, the speech recognition device according to the present invention includes a speech analysis stage-1 for analyzing the speech input to obtain its characteristic parameters, and a feature parameter of the first storage means or the second storage means according to the output of the noise detection means. 3. A speech recognition means for recognizing the input speech by comparing it with the feature parameters determined by the speech analysis means.

[Function 1] In this invention, the background noise is determined and judged to detect its magnitude, and depending on the magnitude, the speaker is given a voice input in a normal voice or a voice input in a loud voice. Instructions 1
1. When recognizing the speaker's voice input, the first storage means or the second storage means is used for selection. For example, when it is determined that the noise is small, the feature parameters stored in the first storage means are used.1. recognize the input voice. Furthermore, when it is determined that the noise is large, the feature parameters stored in the second storage means are used to recognize the human voice.

〔Example〕

FIG. 1 is a block diagram showing the configuration of a speech recognition device according to an embodiment V of the present invention. In the figure, (101) is a microphone, and (102) is a preprocessing means, which amplifies the audio input into the microphone (toi), performs appropriate filter processing, and then converts it into a digital signal. Output to the following digital circuit in the form of 1 data. > (104) is a sound μ (analyzing means) which performs appropriate analysis processing on the voice input through the microphone A-n (101) and pre-processing means (102) to analyze the features necessary for recognition. Convert to parameter.

(105) is a memory in which a standard pattern of characteristic parameters of voice generated (2,
G) is a memory in which the characteristic parameters of the voice generated in a loud voice are stored.
(105) , (10G) l;t: Dictionary t i7) H function. Memory (105), (,10B
) is stored in the voice analysis means (1
04) are respectively performed via the control line (107), and when storing them in the memory (105), the characteristic parameters of the normal loudness of the sound δ are registered in a relatively quiet environment. When storing in the memory (10B), characteristic parameters of voices uttered in a loud voice (7) are registered, but at this time, it is assumed that a loud voice in a quiet environment may sound unnatural. In such a case, a pseudo noise is created using an appropriate noise source 17, and a loud voice is uttered and registered.

(108) is a noise detection means, which has a micrometer size ゛/(
101), and outputs a signal indicating that the noise is small or large depending on whether or not i is larger than a predetermined value. 009) is a changeover switch, which is switched by the output of the noise detection means (108). If the output of the noise detection means (10B) is such that the noise is small, the terminal 010
) and (Ill) are connected. 7 Furthermore, when the output of the noise detection means (ios) is an output indicating that the noise is large, the terminals (1i0) and (112> are connected.

(113) is a speech synthesis means, and a speech analysis means (10
4) and the feature parameters from the memory (105) or (
(115) is a speech synthesis means which calculates the similarity with all the feature parameters of 10G) and outputs the feature parameter with the highest degree of similarity (similarity) to the output terminal (114) as a recognition result. The voice is synthesized (2) according to instructions from the noise detection means (108) and the voice recognition means (113).
, output through the speaker (110).

FIG. 4 is a flowchart showing the operation of the speech recognition device of FIG. 1. When audio 22 starts (S4f11),
First, background noise (or environmental noise) is input via the input microphone (1) (8402).

The noise generated manually is calculated by the noise level in the first stage of noise detection (■8), and if it is determined that the noise is larger than the predetermined threshold, the changeover switch + (10G) is applied.
(7) Connect the ends F (110) and (112) and perform the processes (S407) to (S409) described later. Also, if the noise level is lower than a predetermined threshold and it is determined that the noise is small, the changeover switch is turned off (109).
terminal (110) and cell) are connected, and the process (S404) = (8406) to be described later is executed (S403).

First, the noise detection means (105) detects that i′ is small.
! 1 disconnection 1+, select the changeover switch (109
) terminals (110) and (lit) are connected, and
The voice synthesis means (115) is given an instruction ljλ, and the speaker (
116) to the speaker. and instructs the human voice to "Speak 1. Please" (S404). The voice uttered according to the instructions is inputted to the preprocessing means (102) via the microphone (101) (7), and is then input to the preprocessing means (102).
ζ1. ! After being amplified and further noise removed, A/
It is D-converted and converted into a digital signal. The voice analysis means (104) analyzes the input voice converted into a digital signal and derives characteristic parameters (S405)c.

Next, the voice recognition means <113> performs the voice analysis means (10
4) and memory (10
5), and calculates the similarity with all the feature parameters of the signpost stored in 5), and obtains the pattern with the highest degree of similarity as a recognition result (S4DJ, outputs L( to the output terminal (114))
S4i1), and the process ends (S412).

Next, if the noise detection means (105) determines that the noise is large, the terminals (1i0) and (112> of the changeover switch (109) are connected, and an instruction is given to the speech synthesis means (115). Then, the speaker (11B) is activated and the speaker is instructed to input voice by saying, "Speak loudly" (S407).The voice uttered according to the instructions is transmitted to the microphone (101). Input 17 to the preprocessing means (102) via the preprocessing means (102)
Then, the signal is amplified, noise is removed, and then A/D converted to a digital signal.

The voice analysis means (104,'') analyzes the voice input converted into a digital signal and extracts characteristic parameters (8
408).

Next, the voice recognition means (113) performs the voice analysis means (10
The feature parameters analyzed by 4) and the memory (10
13), and obtains the pattern with the highest degree of similarity as the recognition result (S409), and outputs it to the output end (114) in the same way as in the above case. ) is output L (8411,), and the process ends (8412).

In addition, before outputting the recognition result to the output terminal (114), the voice recognized by the voice recognition means (tia) is synthesized by the voice synthesis means ([15) and output to the speaker (116).
) for the speaker to confirm.

In addition, when instructing the speaker to input voice after determining the level of noise, the CRT (not shown) should be displayed in addition to the speaker (11, G'). It is also possible to have voice commands given to the user.

[Effects of the Invention] As described above, according to the present invention, the volume of the voice is instructed to the speaker according to the magnitude of the background noise, and two types of storage means ( Since the registration of standard feature patterns for normal voices and the registration of feature patterns for loud voices are used, a good recognition rate can be expected in both low-noise environments and high-noise environments.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the structure of a speech recognition device according to an embodiment of the present invention, FIG. 2 is a block diagram showing the structure of a conventional speech recognition device, and FIG. 3 shows the operation of the device shown in FIG. FIG. 4 is a flowchart showing the operation of the apparatus shown in FIG. In the figure, (104) is a voice analysis means, (105) is a memory in which standard patterns of normal voices are stored, (
II) Ili) is a memory in which a standard pattern of loud voices is stored, (108) is a noise detection means, (109)
is a changeover switch, (113) is a voice recognition means, (1
15) is a speech synthesis means.

Claims (1)

[Claims] A first storage means storing feature parameters of normal speech, a second storage means storing feature parameters of loud speech, and measuring and determining background noise to determine its magnitude. a noise level detection means for detecting; a speech synthesis device for instructing the speaker whether to input voice in a normal voice or in a loud voice according to the output of the noise level detection means; and analyze the voice input. speech analysis means for determining the characteristic parameters thereof; feature parameters stored in the first storage means or second storage means according to the output of the noise level detection means; and the characteristic parameters determined by the voice analysis means. A speech recognition device comprising speech recognition means for comparing and recognizing input speech.