JPS60217395A - Critical state detector by voice - 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 [Technical Field of the Invention] The present invention relates to an apparatus for detecting a crisis state that a human being is trying to express by using the voice generated by the human being.

[Technical background of the invention and its problems] In recent years, advances in science and technology such as electronics, mechatronics, large-scale system technology, and robot technology have been remarkable, and with the introduction of factory automation, office automation, laboratory automation, etc. Industrial division has led to rationalization and improved operational efficiency. Furthermore, more advanced home automation systems are being introduced.

On the other hand, pattern recognition technology is also steadily progressing, such as postal code reading devices, alphanumeric characters, j-kana, and even kanji! ! ! l! Optical character recognition devices and online character recognition devices have been put into practical use. moreover,
In the field of speech recognition, word speech recognition devices, continuous speech recognition devices, monosyllable recognition devices, and speech word processors based on Japanese language processing technology have also been put into practical use for specific speakers.

Word speech recognition devices for non-specific speakers have also been put into practical use.
It is used by many users for telephone services such as balance introductions at banks, etc. Furthermore, at the research level, research on continuous speech recognition and speaker recognition is actively being conducted, and with the progress of such research and advances in semiconductor technology and LSI technology, speech recognition devices are being used in a wide range of fields. There is. This invention utilizes the development of such voice recognition technology.

By the way, due to the advancement of science and technology as mentioned above, rationalization and
While efficiency has been achieved, as systems become more complex and larger, humans increasingly encounter various dangers that are different from those in the past.

For example, if we look at robot technology, which has progressed rapidly recently,
Robots are controlled by computers and perform extremely precise movements9, helping to reduce the burden on humans in many aspects such as assembling and transporting parts. However, since robots are not humans, they may malfunction due to electrical noise or programming errors.

As a result of such malfunctions, the current situation is that items may be destroyed or people may be harmed.

Furthermore, not only robots but also large-scale systems such as various plants and power plants are reducing personnel in order to improve efficiency. Therefore, how to accurately and quickly detect abnormal conditions such as earthquakes with a limited number of personnel has become a major issue.

Among humans' means of transmitting information, voice is the most basic and quickest. Therefore, humans can recognize a crisis situation,
When you feel a sense of crisis, speak out before making gestures. For example, if you are a Japanese person, you might say things like "Danger!" or "Stop!"
” “Stop!” “Help!” “Ah!” “Wait!”
"Cats!" etc. Therefore, it would be convenient if voice recognition technology could detect a crisis state by analyzing the above-mentioned utterances, but conventional voice recognition technology, as mentioned above, only uses the linguistic content to detect what the speaker is trying to express. There was a problem in that it was not possible to detect a crisis state.

[Object of the Invention] An object of the present invention is to provide a device that detects when a human expresses a crisis state by voice.

[Summary of the Invention] This invention recognizes the linguistic content of a voice uttered by a human to notify a state of crisis using ordinary voice recognition technology, and at the same time recognizes the absolute level of voice utterance.

It extracts prosodic features such as accent and intonation, recognizes emotions, and uses these to accurately detect the crisis state that a person is trying to express.

That is, the detection device according to the present invention recognizes the linguistic content of the speech by comparing the means for analyzing the input speech and the feature vector of the speech obtained from the analysis result with a speech recognition dictionary created in advance. and a means for detecting the occurrence of a crisis state expressed by a speaker from the linguistic content of the speech recognized by this means and the prosodic features of the speech obtained from the analysis results. It is characterized by

[Effects of the Invention] According to the present invention, prosodic features are used to detect a crisis state, so malfunctions due to speech that appear during normal conversation are reduced and reliability is high, and even if the linguistic content is unclear, This method has the advantage that a crisis state can be reliably detected based on the change pattern of the loudness and prosody of the utterance.

Furthermore, if a speaker verification function is added and the detection results of a crisis state based on the linguistic content and prosodic features described above are used together with the speaker verification results to determine the presence or absence of a crisis state, malfunctions caused by other people's voices can be made. This can further improve reliability.

When the detection device of this invention is used, for example, when a robot is about to harm a human, or when an emergency situation is required due to a failure of the robot, a crisis situation can be automatically detected based on the voice uttered by a worker. This makes it possible to quickly and automatically take action in response to a crisis situation, greatly contributing to improving the safety of various systems and ensuring human safety. .

[Embodiment of the Invention] FIG. 1 shows a schematic configuration of a detection device according to an embodiment of the invention.

In the figure, voice inputted by a microphone (not shown) is inputted to the analysis section 1 as an electrical signal via an amplifier or the like as appropriate. In the analysis section 1, the input audio signal is band-limited to 0 to 5.6 kl-1z using a low-pass filter, and then converted into a 12-bit digital signal at a sampling frequency of, for example, 12 kHz using an A/D converter. Then, frequency analysis is performed using a 16-channel digital bandpass filter, squared, smoothed with a time constant of 16 ms, and a 16-dimensional power spectrum is extracted every 6 ms. Output to 3.

In addition, the analysis section 1 further performs the frequency analysis described above, and also performs, for example, "Speech Recognition" by Yasunaga Niimi, pp. 56-59.
: The fundamental frequency (pitch period) of the speech waveform is extracted by a technique such as the correlation method described in Kyoritsu Shuppan in 1974, and this is output to the prosody information processing section 3 every 8 mS. Furthermore, energy level information is also given to the prosody information processing unit 3 every 8 mS.

The word N recognition unit 2 recognizes the linguistic content of the voice uttered by the speaker.
gil, and sends the recognition result to the crisis state detection section 4. It is effective to configure the language recognition unit 2 to perform recognition using phonemes such as phonemes as basic units, as shown in FIG. 2, for example. In FIG. 2, the phoneme recognition unit 5 uses a phoneme recognition unit 5 to combine a phoneme feature vector consisting of 16×5=80-dimensional vectors for multiple frames of 16-dimensional vectors obtained as a result of analysis of input speech, for example, 5 frames, and a phoneme dictionary 7. A similarity value for each syllable is obtained by matching using the degree method or the like.

Then, the input speech is divided into phonemes using this similarity value, and a phoneme symbol sequence is output to the word recognition unit 6.

The word recognition unit 6 compares the word dictionary 8 and the input phoneme sequence by dynamic programming using a phoneme fusion matrix to obtain a word recognition result.

If phonemes are used as the basic unit of language recognition in this way, there is no need to utter in advance the linguistic content of sounds that can be uttered by humans in a crisis state in order to register them. It has the advantage of being able to respond to speech. In addition, the language recognition unit 2 can be configured to perform not only word recognition but also continuous word recognition as described above, and furthermore, it can be configured to perform word support botting technology that focuses only on word ideas that express crisis. It can also be configured to use

On the other hand, the prosodic information processing unit 3 in FIG. 1 extracts prosodic features related to the expression of a crisis state contained in human speech from the filter output from the analysis unit 1 and the time series of fundamental frequency and energy. FIG. 3 shows an example of the configuration of the prosody information processing section 3. Using the time series of energy and pitch frequency (fundamental frequency) output from the analysis unit 1, speech level is measured and accent and intonation are analyzed.

First, the speech level measurement section 9 measures the maximum speech level of the input voice and outputs it to the crisis state detection section 4 .

9- Also, the accent kernel extraction section 10 extracts the presence of an accent kernel from the pitch pattern and outputs it to the crisis state detection section 4. The intonation determining section 11 determines whether or not the pitch pattern can be uttered in a crisis state, and outputs its certainty, that is, the likelihood, to the crisis state detecting section 4 .

One of the features of the present invention is that the above-mentioned prosodic information processing section 3 is provided, and the crisis state (sense of crisis) expressed by a human is focused on the level of speech and the nature of the pitch pattern of human speech during a crisis. This makes it possible to perform highly accurate detection.

The crisis state detection circuit 4 in FIG. Comprehensive judgment is made by comparing the knowledge regarding the detection of a crisis state, and a signal indicating whether or not there is a crisis state is output.

FIG. 4 shows an example of the configuration of the crisis state detection circuit 4.

Word recognition results, such as "Danger!", "Ah!" 10- "Wait!", etc., information on the linguistic content, information on the maximum energy level, the position of the accent core, intonation tendency, etc. are evaluated by the comprehensive judgment unit 12. It is checked against the state detection knowledge dictionary 13 to determine whether or not there is a crisis state. For example, when it is recognized as "dangerous!", the knowledge in the knowledge dictionary 13 is filled in to represent the crisis state even if the utterance level is low or the accent or intonation is slightly different. In addition, when recognizing "Ah!", the knowledge dictionary 1 determines that it is a crisis state only when the level is high or the pitch pattern expresses a sense of crisis.
3 should be created. In other words, the linguistic content and emotional content are comprehensively judged to determine whether or not there is a crisis state.

FIG. 5 shows another embodiment of the present invention, in which high performance is achieved for a specific speaker by combining the crisis state detection function explained in the embodiment above with speaker verification technology. This is a device for detecting critical situations.

That is, in this embodiment, the speaker matching feature vector obtained from the analysis result of the input speech by the analysis unit 1 and the speaker matching dictionary 15 are matched by the speaker matching unit 14, and the speaker matching result is used for comprehensive judgment. Send to Department 16.

The comprehensive determination unit 16 comprehensively determines the information regarding the crisis state outputted from the crisis state detection unit 4 and the above-mentioned speaker verification results to ultimately determine the presence or absence of the crisis state.

With this configuration, it is possible to prevent false detection of a crisis state by a person other than a specific user, and further improve reliability.

As explained above, according to the present invention, when a person expresses a sense of crisis vocally, the information contained in voice, which is the most natural and quickest means of transmitting information for humans, is used to enhance the sense of crisis. Accurate and stable detection is possible. Therefore,
If the detection device of this invention is applied to a robot, for example, if a human is about to harm the robot, the human can control the robot's movement by vocalizing to stop the robot or make it take an evasive action. Become. Furthermore, when a worker or the like recognizes a crisis state in a large-scale system such as a plant, it becomes possible to automatically take measures to deal with it simply by uttering a voice.

It should be noted that the present invention is not limited to the above-mentioned embodiments, and the specific structure of each block can be modified in various ways without departing from the scope of the invention.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a voice-based crisis detection device according to an embodiment of the present invention, FIG. 2 is a diagram showing an example of the configuration of a language recognition section, and FIG. 3 is an example of the configuration of a prosodic information processing section. A diagram showing
FIG. 4 is a diagram showing an example of the configuration of the crisis state detection section, and FIG. 5 is a diagram showing the configuration of another embodiment of the present invention. DESCRIPTION OF SYMBOLS 1... Analysis section, 2... Language recognition section, 3... Prosodic information processing section, 4... Crisis state detection section, 14... Speaker matching section. Applicant's agent Patent attorney Takehiko Suzue 13-

Claims (1)

[Claims] (1) A means for analyzing input speech, and comparing the feature vector of the speech obtained from the analysis result with a speech recognition dictionary created in advance, and recognizing the linguistic content of the speech. and means for detecting the occurrence of a crisis state expressed by a speaker from the linguistic content of the speech recognized by this means and the prosodic features of the speech obtained from the analysis result. A crisis state detection device using voice. (21) Detection and placement of a crisis state by voice according to claim 1, wherein the detection result of the means for detecting the crisis state expressed by the speaker is used to control the robot 1. (3) A patent claim characterized in that the detection result of the means for detecting a crisis state expressed by a speaker is used together with the speaker verification result to ultimately determine the presence or absence of a crisis state. A device for detecting a crisis state by sound as described in Scope 1.