JP3846500B2 - Speech recognition dialogue apparatus and speech recognition dialogue processing method - Google Patents

Description

  The present invention relates to a voice recognition dialogue apparatus and a voice recognition dialogue processing method for recognizing a voice and performing a response or a specific action corresponding to the recognition result.

  In this type of speech recognition device, there are a specific speaker speech recognition device capable of recognizing only the speech of a specific speaker and an unspecified speaker speech recognition device capable of recognizing the speech of an unspecified speaker.

  The specific speaker voice recognition apparatus registers a standard voice signal pattern of a specific speaker by inputting words that can be recognized by a specific speaker one by one according to a predetermined procedure. After the end, when a specific speaker speaks a registered word, the feature pattern obtained by analyzing the input speech is compared with the registered feature pattern to perform speech recognition. An example of this type of voice recognition dialogue apparatus is a voice recognition toy. For example, about 10 words such as “good morning”, “good night”, “konnichiwa”, etc. are registered in advance as a plurality of command words used as voice commands, and a child who uses the toy registers in advance. ”Compares the audio signal with the registered“ Good Morning ”audio signal, and when both audio signals match, outputs a predetermined electrical signal for the audio command, and to the toy based on this A specific operation is performed.

  Such a specific speaker voice recognition apparatus recognizes only a voice having a specific speaker or a voice pattern close thereto, and as an initial setting, it is necessary to register all the words to be recognized one by one. Was extremely troublesome.

  On the other hand, the unspecified speaker speech recognition apparatus creates standard speech feature data of recognition target words as described above using speech uttered by a large number of speakers (for example, about 200 speakers). It is memorized (registered) so that the speech produced by an unspecified speaker can be recognized with respect to the recognizable words registered in advance.

  In such a speech recognition apparatus, even if a specific speaker speech recognition or unspecified speech recognition is used as speech recognition means, recognition operation and response operation can be performed only after the switch is turned on. It is common.

  However, depending on the type of voice recognition device, if the device is always waiting for voice input and the user can detect voice input and turn on the recognition without turning on the switch each time, It will be highly convenient. For example, in a toy such as a stuffed toy that uses voice recognition, if it is in a voice input standby state, that is, a so-called sleep mode from the normal time, the child can respond immediately by calling the name of the stuffed toy. As a toy, it will be extremely complete. The same applies to all electronic devices that use voice recognition as well as toys.

  As described above, the sleep mode is normally set in the sleep mode, and as a problem when the input voice is detected and the recognition operation is performed, how to suppress the power consumption or whether the input sound signal should be recognized or noise Such that only the voice to be recognized is operated. In particular, in the case of toys and the like, most of them operate with dry batteries, so how to reduce battery consumption is a big issue, and the price of the product needs to be as low as possible. For this reason, the current situation is that there are very many technical restrictions for practical use.

  The present invention has been made in order to solve these problems. In the normal state, only when there is a voice input that can be recognized as a sleep mode, it can detect it and perform a recognition operation. It is intended to minimize power consumption at the time, to enable recognition of speech to be recognized with high accuracy even when noise is present in the usage environment, and to enable practical use at low cost.

  The speech recognition dialogue apparatus of the present invention includes a sound signal input means for inputting a sound signal, a sound signal analysis section for analyzing the sound signal input by the sound signal input means and generating feature data, and registered recognition Standard speech feature data storage means for storing standard speech feature data of possible words and feature data from the sound signal analysis means are input and compared with the stored contents of the standard speech feature data storage means. If it is a speech of a recognizable word, the word detection means for outputting word detection data for the recognizable word, and the word detection data from the word detection means are received to understand the meaning of the input speech , Voice understanding conversation control means for determining response contents corresponding to the meaning of the voice and performing various controls, and voices based on the response contents determined by the voice understanding conversation control means A voice synthesizing means for generating a synthesized output; a voice output means for outputting the voice synthesis output from the voice synthesizing means; and a power of the input sound signal is detected based on a sound signal waveform from the sound signal input means. Input sound signal power detecting means, and judging a noise level based on the power signal from the input sound signal power detecting means, and outputting a response content at a sound level corresponding to the magnitude of the noise level. A speech recognition interactive apparatus, wherein power detection by the input sound signal power detection means is performed for a predetermined time based on a set threshold value, and an input sound signal having a level larger than the threshold value is continuously generated for a predetermined time. And the input sound signal is determined not to be a recognizable word sound, the input sound signal is determined to be noise present in the environment, and the threshold value is set. Updated to a value greater than the level of The noise, the process of determining the noise level, is characterized in that it is carried out in the basis of the magnitude of the threshold value in time of the determination.

  Further, it is characterized in that a response content is output at a sound level corresponding to the power of the input sound signal based on the power signal from the input sound signal power detecting means.

  The speech recognition dialogue processing method of the present invention analyzes the sound signal input by the sound signal input means to generate feature data, and stores the feature data and the standard speech feature data of the recognizable word registered. If the input sound signal is a speech of a recognizable word, the word detection data for the recognizable word is output, and the word detection data is received and input. A speech recognition response output generating step of understanding the meaning of the speech, determining a response content corresponding to the meaning of the speech, and generating a speech synthesis output based on the determined response content, and the sound signal input means An input sound signal power detecting step for detecting the power of the input sound signal based on the sound signal waveform from the input signal, and determining a noise level based on the power signal from the input sound signal power detecting means. A speech recognition dialogue processing method for outputting response contents at a voice level corresponding to the magnitude of the noise level, wherein power detection in the input sound signal power detection step is performed for a predetermined time with reference to a set threshold value. If the input sound signal having a level greater than the threshold value is continuously present for a predetermined time and it is determined that the input sound signal is not a recognizable word sound, the input sound signal is in the environment. The threshold value is updated to a value larger than the noise level, and the noise level is determined based on the threshold value at the time of the determination. It is characterized by that.

  In addition, based on the power signal from the input sound signal power detection step, response contents are output at a sound level corresponding to the power of the input sound signal.

  As described above, according to the present invention, the noise level is determined based on the power signal from the input sound signal power detection means, and the response content is output according to the sound level corresponding to the magnitude of the noise level. . Therefore, when the noise level is high, response content output with a large voice level can be generated, and the response content can be easily heard even if some noise exists in the environment.

  Then, the threshold value is updated to a value larger than the steady noise level, and the noise level at that point is determined based on the threshold value at a certain point in time. The magnitude index can be obtained from the threshold value, and the current noise level can be easily determined. Moreover, even if the noise level fluctuates, it is possible to generate a response output with a voice level corresponding to the fluctuation, and it is possible to obtain a response output with a voice level that is more adapted to noise in the environment.

  In addition, response contents are output according to the sound level corresponding to the power of the input sound signal, so that when the speaker asks a loud voice, the speaker responds with a loud voice, and when the speaker asks a small voice, the response level is small. By performing an operation of responding with a voice, it is possible to perform a dialogue at a voice level according to the situation.

  Embodiments of the present invention will be described below with reference to the drawings. In this embodiment, a case where the present invention is applied to a toy is taken as an example, and here, a case where the present invention is applied to a stuffed animal such as a dog will be described as a toy for an infant. An example in which the present invention is applied to an unspecified speaker voice recognition apparatus capable of recognizing an unspecified speaker's voice will be described.

(First embodiment)
FIG. 1 is a block diagram for explaining a first embodiment of the present invention. FIG. 1 is a block diagram illustrating an embodiment of the present invention. In general, a sound signal input unit 1, a sound signal analysis unit 2, a word detection unit 3, a standard voice feature data storage unit 4, a voice understanding conversation control Unit 5, response data storage unit 6, speech synthesis unit 7, speech output unit 8, input sound signal power detection unit 9, and the like. Of these components, the sound signal analysis unit 2, the word detection unit 3, the standard speech feature data storage unit 4, the speech understanding conversation control unit 5, the response data storage unit 6, the speech synthesis unit 7, etc. The stuffed animal is housed in the abdomen, the sound signal input unit (microphone) 1 is provided in the stuffed ear, for example, and the voice output unit (speaker) 7 is provided in the mouth, for example. The functions of these units will be described in order below.

  The sound signal input unit 1 inputs a sound signal (including noise and the like) such as a voice spoken by a speaker, and is composed of a microphone, an amplifier, a low-pass filter, an A / D converter, and the like (not shown). The sound signal input from the microphone is converted into an appropriate sound waveform through an amplifier and a low-pass filter, then converted to a digital signal (for example, 12 KHz.16 bits) by an A / D converter, and the output is the sound signal. The data is sent to the analysis unit 12. The sound signal analysis unit 12 performs frequency analysis on the waveform signal sent from the sound signal input unit 1 every short time using a computing unit (CPU), and extracts a several-dimensional feature vector representing the frequency feature. (LPC-CEPSTRUM coefficient is generally used), and a time series of this feature vector (hereinafter referred to as a feature vector sequence) is output.

  The standard voice feature data storage unit 4 stores (registers) standard patterns of recognition target words created in advance using voices spoken by a large number of speakers (for example, about 200 speakers) for each word. ) ROM. In this case, because it is a stuffed toy as an example, the recognition target word is set to about 10 words, as is the word, for example, "Good morning", "good night", "Hello", "tomorrow", used to greeting such as "weather" There are many words, but it is not limited to this. Various words can be registered, and the number of registered words is not limited to ten.

  Although not shown, the word detection unit 3 is mainly composed of an arithmetic unit (CPU) and a ROM storing a processing program, and a word registered in the standard voice feature data storage unit 4 is included in the input voice. This is to detect the certainty in which part of the body. As this word detection unit 3, a hidden Markov model (HMM) method, a DP matching method, or the like can be used. However, here, a keyword spotting processing technology based on a DRNN (dynamic recurrent neural network) method (for this technology, The present applicant has already applied for a patent according to Japanese Patent Laid-Open Nos. 6-4097 and 6-119476.), And word detection for enabling speech recognition close to continuous speech recognition by an unspecified speaker. It is assumed that data is output.

  Specific processing of the word detection unit 3 will be briefly described with reference to FIG. The word detection unit 3 has a certain degree of certainty in which part of the input sound signal (in this case, the input sound signal is a speech) the word registered in the standard sound feature data storage unit 4 exists. Is detected. Now, it is assumed that a voice such as “Tomorrow's weather is ...” is input from the speaker, and a voice signal as shown in FIG. Among the phrases “Tomorrow's weather is ...”, “Tomorrow” and “Weather” are keywords in this case, and these are standard words as one of about 10 registered words registered in advance. The pattern is stored in the voice feature data storage unit 4. When these registered words are, for example, 10 words, a signal for detecting each word is output corresponding to these 10 words (referred to as word 1, word 2, word 3,...). Therefore, from the information such as the value of the detection signal, it is detected to what extent the corresponding word exists in the input speech. That is, when the word “weather” (word 1) is present in the input voice, the detection signal waiting for the signal “weather” is “weather” of the input voice as shown in FIG. Stand up at the part. Similarly, when the word “Tomorrow” (word 2) is present in the input voice, the detection signal waiting for the signal “Tomorrow” is “ Get up in the “Tomorrow” part. In FIGS. 5B and 5C, a numerical value such as 0.9 or 0.8 is a numerical value indicating the probability (approximation degree). If the numerical value is high such as 0.9 or 0.8, the registered word having the high probability is It can be said that the input speech is a recognition candidate. That is, the registered word “tomorrow” is present with a certainty of 0.8 in the w1 portion on the time axis of the input audio signal, as shown in FIG. As shown in FIG. 5B, it can be seen that the input audio signal exists with a certainty of 0.9 in the portion w2 on the time axis.

  In the example of FIG. 2, in response to an input of “weather”, a signal waiting for the word 3 (assuming that this word 3 is a registered word “what time”) as shown in FIG. In the w2 portion on the time axis, it stands up with a certain degree of certainty (the value is about 0.6). Thus, when two or more registered words exist as recognition candidates at the same time with respect to the input speech signal, the word having the highest degree of approximation (a numerical value indicating the probability) is used as the recognition word. A method for selecting, a correlation table showing a correlation rule between each word is prepared in advance, and a method for selecting any one word as a recognition word by using this correlation table is used. Determine the word. For example, if the recognition method is determined by the former method, in this case, the degree of approximation corresponding to the w2 portion on the time axis is the highest degree of approximation of the detection signal for detecting “weather”. It is determined that the recognition candidate for the partial input speech is “weather”. Note that the speech understanding conversation control unit 5 recognizes the input speech based on these approximations.

  The voice comprehension / conversation control unit 5 is mainly composed of an arithmetic unit (CPU) and a ROM storing a processing program. The word detection data from the word detection unit 3 is input and the word detection data is used as a basis. , Recognizes the voice (understands the meaning of the entire input voice), determines the response content according to the meaning of the input voice by referring to the response data storage unit 6 composed of ROM, and synthesizes the voice The output signal is sent to the unit (mainly composed of CPU and ROM) 7 and the audio output unit 8. For example, detection data from the word detection unit 3 as shown in FIGS. 2B to 2E (this is called a word lattice. This word lattice is a registered word name, a degree of approximation, a word start point s and an end point e. 1 or the like is included, first, one or more words as keywords in the input speech are determined based on the word lattice. In this example, since the input voice is “Tomorrow's weather is ...”, “Tomorrow” and “Weather” will be detected, and “Tomorrow” will be detected from the keywords “Tomorrow” and “Weather”. Understand the contents of the continuous voice saying "The weather is ..." and select and output the response contents corresponding to it. In this case, the response content is a response content such as “Tomorrow's weather is sunny”. This is a state detection means (temperature detection unit, atmospheric pressure detection unit, calendar, which is not shown here). For example, if the information is related to the weather, the change in the weather is judged based on the change in the atmospheric pressure from the atmospheric pressure detection unit, and if the atmospheric pressure tends to rise, it will be dealt with The response content thus read is read out from the response data storage unit 6. Similarly, responses relating to temperature, time, date, etc. are possible.

  Further, the speech recognition process close to continuous speech recognition by the keyword spotting process described above can be applied not only to Japanese but also to other languages. For example, if the language used is English, the registered recognizable words are, for example, “good morning”, “time”, “tommorow”, “good night”, etc. Feature data of recognizable registered words is stored in the standard voice feature data storage unit 4. And now, when the speaker asks "what time is it now", the word "time" becomes the keyword in this case, and the word "time" is entered. The detection signal waiting for the “time” audio signal rises at the “time” portion of the input audio. When the detection data (word lattice) from the word detection unit 3 is input, first, based on the word lattice, one or more words as keywords in the input speech are determined. In this example, since the input voice is “what time is it now”, “time” is detected as a keyword, and “what time is it now” is continuously generated based on this keyword. Understand the content of sound.

  Note that a CPU for controlling the above-described speech analysis, word detection, speech understanding conversation control, speech synthesis, and the like may be provided, but here, as shown in FIG. It is assumed that the main CPU 10 is provided and the entire processing of the present invention is performed by this single CPU 10.

  The above is an example of recognition and response operations for input voice (speaker's spoken voice). In such a voice recognition dialogue apparatus, in the first embodiment of the present invention, the sleep mode is normally set. When the voice to be recognized is input, the input voice is detected and recognized. The first embodiment will be described below.

  The input sound signal power detection unit 9 shown in FIG. 1 detects the power of the input sound signal (including noise and the like) output from the sound signal input unit 1. For example, in FIG. The power signal is detected from the input sound signal waveform as shown in FIG. The input sound signal power detection unit 9 is in a state where it can always detect the power of the input sound signal captured by the sound signal input unit 1 regardless of whether the apparatus is in the sleep mode or the operation mode.

  By the way, as an initial setting means for setting the device to the sleep mode, for example, a means for automatically setting the sleep mode when a stuffed battery is set, a sleep mode setting switch or a sleep mode changeover switch is provided. Various means such as switching to the sleep mode by turning on or switching the changeover switch to the setting mode side can be considered. When the sleep mode is set by any of these means, a sleep mode request is issued to the CPU, the flag is set to the sleep mode side, and the sleep mode is entered. In addition, after the initial setting of the sleep mode, to enter the sleep mode after the conversation with the stuffed animal, if there is no question from the speaker for a certain period of time, or if the speaker "Sleep" or "good night" is determined as a sleep mode request on the device side, and as described above, the flag is set to the sleep mode side to set the sleep mode.

  In such a configuration, the process will be described with reference to the flowchart of FIG.

  First, it is assumed that the initial setting of the sleep mode has already been made. Even in the sleep mode, the input sound signal power detection unit 9 can always detect the power of the sound signal captured by the sound signal input unit 1, but at this point, the CPU 10 of the apparatus is in a non-operating state. .

  Now, it is assumed that the sound signal input unit 1 inputs some sound signal, and the sound signal waveform as shown in FIG. Thereby, the input sound signal power detection unit 9 performs the input sound signal power detection process as follows.

  That is, the input sound signal power detection means 9 determines whether or not the power of the input sound signal is greater than a preset threshold value th1 (step s1), and the power of the input sound signal is greater than the threshold value th1. At that time, a signal indicating that a sound signal has been input is output to the CPU 10 (time t1 in FIG. 3B). Thereby, the CPU 10 rises (step s2), and after a predetermined time (about 10 msec) from time t1, the voice analysis unit 2 performs a voice analysis process based on the signal from the sound signal input unit 1 (step s3). As can be seen from FIG. 3B, since it takes about 20 msec from the time t0 when the sound signal is input to the start of the sound signal analysis process, the actual sound signal analysis process is performed. Is started 20 seconds after the sound signal is input, but it does not particularly affect the word recognition process.

  Then, after performing the word detection process (step s4) by comparing with the standard voice feature data using the sound signal analysis result, it is determined whether the apparatus is currently in the sleep mode or the operation mode (step s5). This sleep mode or operation mode is determined by looking at the flag. In this case, since the flag is set to the sleep mode, the input sound signal is registered in advance using the word detection data. It is determined whether the keyword is a recognizable keyword (step s6). In this determination, if it is determined that the input sound signal is not a keyword, the input sound signal is regarded as noise or some kind of sound different from the question about the stuffed animal, and a process of lowering the CPU 10 is performed (step s7). Returning to step s1, the state again waits for the sound signal input.

  On the other hand, if it is determined in step s6 that the input sound signal is a keyword sound, a flag for setting the device from the sleep mode to the operation mode is set (step s8). For example, if the input sound signal is the voice of the keyword “Good morning”, the mode of the apparatus is switched to the operation mode for the first time, and the voice understanding conversation control process is performed on the input voice (step s9). Then, it is determined whether or not the input voice has been completed for one utterance (step s10), and if not, the process returns to the sound signal analysis process (step s3). On the other hand, if it is determined that the input speech has been completed for one utterance, the process proceeds to a determination process (step s11) of whether there is a sleep mode request. Specifically, whether or not the input voice has been completed for one utterance is whether or not a certain conversation has ended. For example, when a speaker asks "Good morning", the stuffed toy responds "Good morning, I'm up early today", and the conversation may continue. As described above, while the conversation continues, it is determined that the voice input has not been completed, the sound signal analysis of the content asked by the speaker is performed (step s3), and the word detection process is performed (step s4). In this case, since the operation mode is already set instead of the sleep mode, the speech understanding conversation control process (step s9) is entered directly after the word detection process (step s4). This is a normal speech recognition processing operation, and as described above, the keyword spotting processing is performed to recognize a word as a keyword from speech uttered by a speaker and output the corresponding response content.

  When a series of conversations (one utterance) ends, it is determined whether or not a sleep mode request has been made (step s11). This sleep mode request is determined, for example, when the speaker says "quiet" or "good night" to the stuffed toy, or when there is no voice input for a certain period of time, it is determined that the sleep mode is requested. . If there is no sleep mode request among these sleep mode requests, it is assumed that the conversation will continue, a response output to the speaker's inquiry is generated (step s12), and the processing from step s3 is repeated. On the other hand, if any of the sleep mode requests is issued, a response output is issued as necessary (step s13), and then the flag is set to the sleep mode (step s14). As a result, the apparatus enters the sleep mode (step s15) and returns to the state of step s1 again. In the process of step 13, for example, if the sleep request is “good night”, the flag is set to the sleep mode after responding to “good night” from the stuffed toy.

  As described above, by placing the device (in this case, a stuffed animal) in sleep mode, when a speaker wants to talk to the stuffed animal, the user can be asked by a recognizable word without having to bother to turn on the switch. The stuffed animal responds to it. For example, it is also possible for a stuffed animal to respond in response to the name being called by simply calling the name of the stuffed animal (assumed to be registered) without turning on the switch.

  Moreover, if it is applied not only to toys but also to electronic devices that operate by recognizing sound, it is possible to perform the corresponding operation by simply emitting some predetermined sound without the trouble of turning on the switch. It becomes possible and its application range is very wide.

  In the present invention, in the sleep mode, the CPU 10 is in an inoperative state, and only the sound signal input unit 1 and the input sound signal power detection unit 9 are operable. In this type of device, the power consumption of the CPU 10 is large, but the power consumption of the sound signal input unit 1 and the input sound signal power detection unit 9 is usually very small compared to this. In the case of the present invention, when a sound signal of a certain level or higher is input, the CPU 10 temporarily rises, but even if the input sound signal is a voice, it is determined that the voice is not a recognized word. Then, the CPU 10 immediately falls. The operation mode is not entered until the speech of a recognizable word (keyword) having a power of a certain level or higher is input. Therefore, the power consumption in the sleep mode can be kept extremely low.

  By the way, the input sound signal power detection unit 9 compares the input sound signal power with a certain threshold value, and when the input sound signal power exceeds the threshold value, the input sound signal detection is present in the environment. This is to avoid the influence of noise and the like. This will be described below.

  For example, if there are continuous running noises of automobiles in the vicinity of an expressway or continuous noises during construction, these noises are regarded as having an input sound signal and the CPU 10 is started up. Is not preferred. In order to cope with this, a threshold value is set, and the CPU 10 is started up only when a sound signal exceeding the threshold value is input (the mode is still the sleep mode at this time). At this time, the threshold value can also be updated so as to cope with the fluctuation of the noise level.

  That is, as described above, the sleep mode is initially set when the battery is inserted or when the sleep mode is set by some switch as described above. There is a case where the stationary noise as described above is continuously inputted without the voice being inputted for a while. For example, as shown in FIG. 5 (a), for a while from time t0, an input sound signal waveform A1 due to stationary noise is output, and then, for example, a speech waveform A2 of a recognizable word “Good morning” is output. Suppose. FIG. 4B shows a power signal corresponding to the input sound signal (including sound). At this time, if the initially set threshold value is th1, if the noise level is greater than the threshold value th1, the processing from step s2 to step s6 in the flowchart of FIG. The sound signal analysis is also performed for the noise, and then processing for detecting a word and determining whether or not the keyword is a keyword is performed. However, in this case, since the input sound signal is noise and not the keyword sound, the process from step s1 to step s6 is repeated for a while. If such processing continues for a certain period of time, it is determined that the input sound signal is noise, and the threshold value is reset to a value th2 higher than th1 (see FIG. 5B).

  When a sound signal having a level greater than the newly set threshold value th2 is input and it is determined that the input sound signal is the voice of a recognized word (keyword), the flag operates from the sleep mode. The mode is switched to the mode, and the apparatus enters the operation mode only at that time.

  In this way, by changing the threshold value according to the noise level, it is possible to reliably detect only the speech to be recognized without being affected by stationary noise that continues continuously for a long time.

(Second embodiment)
In the first embodiment described above, the power of the input sound signal is detected to determine the input of the sound signal, and only when the sound signal is input, the CPU 10 is started up once and the input sound signal is the keyword. In the second embodiment, a case has been described in which the operation mode is set when a determination is made as to whether the keyword is a keyword and a series of processing such as recognition processing and response control processing is performed. When there is noise in the environment, processing to set the response level according to the level of the noise level, and detect the power of the input voice and set the response level based on the power level Processing to be performed will be described.

  First, an example in which the response level is set according to the noise level when noise exists in the environment will be described. In order to realize this, the above-described threshold value is used as an index for judging the magnitude of noise. That is, in the first embodiment, the threshold value set so as not to be affected by noise is initially th1, and the threshold value newly set corresponding to the magnitude of the subsequent noise level is th2. (Th1 <th2).

  Since this threshold value is set corresponding to the magnitude of the noise level, the threshold value at that time can be used as an index representing the magnitude of the noise level. That is, the CPU 10 judges which threshold value is currently used, and sets a response output level corresponding to the threshold value in advance, so that the response output level corresponding to the noise level at that time can be used. Response contents can be output.

  For example, if the threshold is now th1, it is determined that the noise level is not so high, and the response content is output at a response output level set in advance accordingly. If the threshold value is th2, it is determined that the noise level has become slightly higher, and response contents are output at a response output level set in advance accordingly. In this way, by setting the response output level according to the level of the noise level present in the environment, it responds with a loud sound when the noise level is somewhat high, so even if there is some noise level in the environment The response is easy to hear.

  Next, an example of setting a response level based on the power level of the input sound signal will be described. This performs an operation of responding with a loud voice when the speaker asks with a loud voice and responding with a small voice when the speaker asks with a small voice. For example, when the surroundings are quiet, such as at midnight, it is desirable that the speaker and the device interact with each other with a low voice so as not to disturb other people. Conversely, in a noisy environment, it is natural to have a conversation at a certain volume. In addition, it is common for a person with a long ear to speak with a loud voice, in which case it is desirable to respond with a loud voice. Based on such points, the level of the input sound signal is detected and a response output level corresponding to the level is set.

  Suppose that a speaker asks in the middle of the night, for example, “What time is it now?” Here, even if the apparatus is in the sleep mode, if the word “what” is a recognized word and the signal level is higher than the threshold value at that time, the CPU 10 enters the operation mode and the recognition response Operation is possible. When the CPU 10 is operating, the power signal of the input sound signal is always input, and the CPU 10 can monitor the power signal and know the power of the input sound signal at the present time. As a result, the power signal of the voice “now” is compared with, for example, a preset threshold value, and the CPU 10 responds according to the input voice power based on the comparison result. Processing such as generating output is performed. That is, if the input voice level is high, response contents are generated according to a response output with a large sound level, and if the input sound level is low, response contents are generated according to a response output with a small sound level. In this case, since the speaker asks “What time is it now” with a small voice, for example, a response content such as “It is now 11:00” is output at a low voice level.

  In this way, when the speaker asks with a loud voice, the speaker responds with a loud voice, and when the speaker asks with a small voice, the speaker responds with a small voice. When is quiet, it is possible to have a conversation at a volume according to the situation at that time so as not to disturb other people.

  If there is noise in the environment described above, the process of setting the response level according to the level of the noise level and the power of the input sound signal are detected and based on the level of the power. The processing for setting the response level can be applied to the device that can set the sleep mode described in the first embodiment, but it can also be applied to the device that cannot set the sleep mode. it can.

  Moreover, although this invention demonstrated in the example applied to the unspecified speaker voice recognition apparatus, it is applicable also to a specific speaker voice recognition apparatus.

  Moreover, although the example which applied this invention to the stuffed toy as a toy was demonstrated in the above each Example, it is not restricted to a stuffed toy. Needless to say, the present invention can be applied to other toys. Furthermore, it can be applied not only to toys but also to game machines and various electronic devices used in daily life, and the application range is considered to be extremely wide.

  As described above, the speech recognition dialogue apparatus generates feature data by analyzing the input sound signal, compares the feature data with the standard speech feature data of recognizable words registered in advance, and inputs the feature data. If the received sound signal is a recognizable word, the word detection data for the word is output, the word detection data is received, the meaning of the input speech is understood, and the corresponding response content is determined. Means for outputting, input sound signal power detecting means for detecting the power of the input sound signal based on the sound signal waveform of the input sound signal, and the apparatus enters the sleep mode in response to a sleep mode request for setting the apparatus to the sleep mode. A process for determining whether the power of the input sound signal is equal to or greater than a predetermined threshold and whether the input sound signal is the sound of the recognizable word. When the power of the input sound signal is equal to or higher than the threshold value and the input sound signal is determined to be the sound of the recognizable word, the apparatus is changed from the sleep mode to the operation mode. Therefore, only when the input sound signal is a voice of a recognized word having a certain level or higher, it is possible to switch from the sleep mode to the operation mode. In this case, since it is not an operation mode, only the speech of words that should be recognized can be targeted for recognition response processing without being affected by noise, etc. In the sleep mode, only the portion with low power consumption operates. Therefore, the power consumption in the sleep mode can be made extremely small. That is, in the present invention, in the sleep mode, only the low power consumption part such as the sound signal input unit and the input sound signal power detection unit is in an operable state, and when a sound signal of a certain level or higher is input, The CPU temporarily rises, but when it is determined that the input sound signal is not the voice of the recognized word, the CPU immediately falls. In addition, since the operation mode is set only when a voice of a recognizable word (keyword) having a certain level or more of power is input, the power consumption in the sleep mode can be extremely low. .

    In addition, the power detection of the input sound signal when switching from the sleep mode to the operation mode is performed for a predetermined time with the power detection in the sleep mode as a reference for the preset threshold value, When it is determined that the input sound signal is continuously present for a predetermined time and the input sound signal is not a recognizable word voice, the input sound signal is determined to be stationary noise existing in the environment. As a result, sound signals that have a certain level in the environment and exist continuously for a long time can be determined to be noise, and should not be affected by such noise. Can do.

  In addition, the power detection of the input sound signal when switching from the sleep mode to the operation mode is performed for a predetermined time with the power detection in the sleep mode as a reference for the preset threshold value, and the power level is higher than the threshold value. When it is determined that the input sound signal is continuously present for a predetermined time and the input sound signal is not recognizable, the input sound signal is determined to be stationary noise existing in the environment. At the same time, since the threshold value is updated to a value larger than the stationary noise level, for a voice having a certain level in the environment and continuously existing for a long time, The sound signal level can be detected and a threshold value can be set according to that level, so that even if the noise in the environment is larger than the threshold value that was initially set, the noise level It is possible to accurately determine, it is possible to prevent the influence of noise, such as is present on a regular basis in the environment.

  In addition, the input sound signal is analyzed to generate feature data, and the feature data is compared with standard speech feature data of a recognizable word registered in advance. If there is, means for outputting word detection data for the word, receiving the word detection data, understanding the meaning of the input speech, determining and outputting the corresponding response content, sound signal of the input sound signal An input sound signal power detecting means for detecting the power of the input sound signal based on the waveform is provided, the noise level in the environment is determined based on the power signal from the input sound signal power detecting means, and the noise level is increased. Since the response contents are output with the sound level corresponding to the sound level, the response contents can be output with a large sound level when the noise level in the environment is large. Can be noise is easier to hear the response content be present.

  Further, as a means for judging the noise level in the environment based on the power signal of the input sound signal, a threshold for noise removal is set, and the power is detected for a predetermined time based on the threshold, If an input sound signal with a level greater than the threshold value exists continuously for a predetermined time and the input sound signal is determined not to be a recognizable word sound, the input sound signal exists in the environment. The threshold value is updated to a value larger than the stationary noise level, and the noise level at that point is determined based on the threshold value at a certain point in time. Thus, the noise level index can be obtained from the threshold value, and the current noise level can be easily determined. Moreover, even if the noise level fluctuates, it is possible to generate a response output with a voice level corresponding to the fluctuation, and it is possible to obtain a response output with a voice level that is more adapted to noise in the environment.

  Also, the input speech is analyzed to generate speech feature data, the speech feature data is compared with the standard speech feature data of recognizable words registered in advance, and the input sound signal can be recognized If it is, means for outputting word detection data for the word, receiving the word detection data, understanding the meaning of the input speech, determining and outputting the corresponding response content, sound of the input sound signal Input sound signal power detecting means for detecting the power of the input sound signal based on the signal waveform is provided, and the response content by the sound level corresponding to the power of the input sound signal based on the power signal from the input sound signal power detecting means Since the output is performed, when the speaker asks with a loud voice, it responds with a loud voice, and when the speaker asks with a small voice, it responds with a low voice. For example, when the surroundings are quiet, such as late at night, the speaker and the device can talk with each other with a low voice so that other people will not be disturbed. It becomes possible.

  In addition, the speech recognition dialogue processing method analyzes the input sound signal to generate feature data, compares the feature data with standard speech feature data of recognizable words registered in advance, and inputs the input sound. If the signal is a recognizable word, outputting word detection data for the word, receiving the word detection data, understanding the meaning of the input speech, determining and outputting the corresponding response content An input sound signal power detection step for detecting the power of the input sound signal based on the sound signal waveform of the input sound signal is provided, and when the device is in the sleep mode due to a sleep mode request for setting the device in the sleep mode. Performing a process of determining whether the power of the input sound signal is equal to or higher than a predetermined threshold and whether the input sound signal is the sound of the recognizable word, When the power of the sound signal is equal to or higher than the threshold value and the input sound signal is determined to be the sound of the recognizable word, the device is changed from the sleep mode to the operation mode. Only when the sound signal is a voice of a recognized word having a certain level or higher, the operation mode can be changed from the sleep mode. When the noise is low or the signal level is high, the word is not a recognized word. Because it is not an operation mode, only words that should be recognized without being affected by noise or the like can be targeted for recognition and response processing, and only the portion with low power consumption is active in sleep mode. The power consumption in the sleep mode can be made extremely small. That is, in the present invention, in the sleep mode, only the low power consumption parts such as the sound signal input unit and the input sound signal power detection unit are operable, and when a sound of a certain level or higher is input, The CPU starts up, but when it is determined that the input sound signal is not a recognized word, the CPU immediately goes down. In addition, since the operation mode is set only when a voice of a recognizable word (keyword) having a certain level or more of power is input, the power consumption in the sleep mode can be extremely low. .

  In addition, the power detection of the input sound signal when switching from the sleep mode to the operation mode is performed for a predetermined time with the power detection in the sleep mode as a reference for the preset threshold value, When it is determined that the input sound signal is continuously present for a predetermined time and the input sound signal is not a recognizable word voice, the input sound signal is determined to be stationary noise existing in the environment. As a result, sound signals that have a certain level in the environment and exist continuously for a long time can be determined to be noise, and should not be affected by such noise. Can do.

  In addition, the power detection of the input sound signal when switching from the sleep mode to the operation mode is performed for a predetermined time with the power detection in the sleep mode as a reference for the preset threshold value, and the power level is higher than the threshold value. When it is determined that the input sound signal exists continuously for a predetermined time and the input sound signal is not a recognizable word, the input sound signal is determined to be stationary noise existing in the environment, and Since the threshold value is updated to a value larger than the steady noise level, for a sound signal having a certain level in the environment and continuously existing for a long time, The level of the sound signal can be detected and a threshold value can be set according to that level, so that even if the noise in the environment is larger than the threshold value that was initially set, the noise is targeted. Can be determined, it is possible to prevent the influence of noise, such as is present on a regular basis in the environment.

  In addition, the input sound signal is analyzed to generate feature data, and the feature data is compared with standard speech feature data of a recognizable word registered in advance. If there is, the step of outputting word detection data for the word, receiving the word detection data, understanding the meaning of the input voice, determining and outputting the corresponding response content, the voice waveform of the input sound signal An input sound signal power detection process for detecting the power of the input sound signal based on the input sound signal is provided. Based on the power signal from the input sound signal power detection process, the noise level in the environment is determined and the magnitude of the noise level is determined. The response contents are output with the sound level corresponding to the sound level. When the environmental noise level is high, the response contents can be output with a large sound level. Even if noise is present can be easier to hear the response content.

  Further, as a means for judging the noise level in the environment based on the power signal of the input sound signal, a threshold for noise removal is set, and the power is detected for a predetermined time based on the threshold, If an input sound signal with a level greater than the threshold value exists continuously for a predetermined time and the input sound signal is determined not to be a recognizable word sound, the input sound signal exists in the environment. The threshold value is updated to a value larger than the stationary noise level, and the noise level at that point is determined based on the threshold value at a certain point in time. Thus, the noise level index can be obtained from the threshold value, and the current noise level can be easily determined. Moreover, even if the noise level fluctuates, it is possible to generate a response output with a voice level corresponding to the fluctuation, and it is possible to obtain a response output with a voice level that is more adapted to noise in the environment.

In addition, the input sound signal is analyzed to generate feature data, and the feature data is compared with standard speech feature data of a recognizable word registered in advance. If there is, the step of outputting word detection data for the word, receiving the word detection data, understanding the meaning of the input speech, determining and outputting the corresponding response content, the sound signal of the input sound signal An input sound signal power detection process for detecting the power of the input sound signal based on the waveform is provided. Based on the power signal from the input sound signal power detection process, response contents are output at a sound level corresponding to the power of the input sound signal. Because when the speaker asks with a loud voice, the speaker responds with a loud voice, and when the speaker asks with a small voice, the speaker responds with a low voice. For example, when the surroundings are quiet, such as late at night, the speaker and the device can talk with each other with a low voice so that other people will not be disturbed. It becomes.

The block diagram explaining the structure of the 1st Example of this invention. The figure explaining the speech recognition process by the word detection process by a word detection part, and the speech understanding conversation control part. The figure which shows an example of an input sound signal waveform and its power signal. The flowchart explaining the process in the Example. The figure explaining the example which sets the threshold value setting and response output level according to the noise level in a 1st, 2nd Example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Sound signal input part 2 ... Sound signal analysis part 3 ... Word detection part 4 ... Standard voice characteristic data storage part 5 ... Voice understanding conversation control part 6 ... Response data storage part DESCRIPTION OF SYMBOLS 7 ... Speech synthesis part 8 ... Voice output part 9 ... Input sound signal power detection part 10 ... CPU
th1, th2 ... threshold

Claims (4)

A sound signal input means for inputting a sound signal;
A sound signal analyzer for analyzing the sound signal input by the sound signal input means and generating feature data;
Standard speech feature data storage means for storing standard speech feature data of registered recognizable words;
Feature data from the sound signal analysis means is input and compared with the stored contents of the standard voice feature data storage means. If the input sound signal is a recognizable word speech, word detection for the recognizable word is performed. Word detection means for outputting data;
Receiving the word detection data from the word detection means, understanding the meaning of the input voice, determining the response content corresponding to the meaning of the voice and performing various controls; voice understanding conversation control means;
Speech synthesis means for generating speech synthesis output based on the response content determined by the voice understanding conversation control means;
Voice output means for outputting the voice synthesis output from the voice synthesis means to the outside;
Input sound signal power detection means for detecting the power of the input sound signal based on the sound signal waveform from the sound signal input means;
Have
Based on the power signal from the input sound signal power detection means, it is a voice recognition dialogue apparatus for judging the noise level and outputting the response content by the voice level corresponding to the magnitude of the noise level,
Power detection by the input sound signal power detection means is performed for a predetermined time with a set threshold as a reference, and an input sound signal having a level higher than the threshold continuously exists for a predetermined time, and the input If it is determined that the sound signal is not a recognizable word speech, the input sound signal is determined to be noise present in the environment, and the threshold is updated to a value greater than the noise level.
A speech recognition dialogue apparatus characterized in that the process of determining the noise level is performed based on a threshold value at the time of the determination.   2. The speech recognition dialogue apparatus according to claim 1, wherein a response content is output at a voice level corresponding to the power of the input sound signal based on a power signal from the input sound signal power detection means. Analyzing the sound signal input by the sound signal input means to generate feature data, comparing the feature data with the standard speech feature data storing the standard speech feature data of the recognized recognizable word, and the input If the received sound signal is a speech of a recognizable word, word detection data for the recognizable word is output, the word detection data is received, the meaning of the input speech is understood, A speech recognition response output generating step of determining a response content corresponding to the meaning and generating a speech synthesis output based on the determined response content;
An input sound signal power detection step for detecting the power of the input sound signal based on the sound signal waveform from the sound signal input means;
Have
A speech recognition dialogue processing method for determining a noise level based on a power signal from the input sound signal power detection means and outputting a response content by a voice level corresponding to the magnitude of the noise level,
The detection of power in the input sound signal power detection step is performed for a predetermined time with a set threshold as a reference, an input sound signal having a level larger than the threshold is continuously present for a predetermined time, and the input If it is determined that the sound signal is not a recognizable word speech, the input sound signal is determined to be noise present in the environment, and the threshold is updated to a value greater than the noise level.
A speech recognition dialogue processing method, wherein the process of determining the noise level is performed based on a threshold value at the time of the determination.   4. The speech recognition dialogue processing method according to claim 3, wherein response contents are output at a voice level corresponding to the power of the input sound signal based on the power signal from the input sound signal power detection step.

Images