JPH08263092A - Response voice generating method and voice interactive system - Google Patents

Abstract

PURPOSE: To improve the degree of understanding and to shorten a response time in a voice interactive system of understanding the contents of an input voice and outputting a proper response based on the contents. CONSTITUTION: When the voice is inputted, a voice recognition part 21 performs instantaneously recognition processing, and sends the recognition result to a language analysis processing part 4. The language analysis processing part 4 outputs sentence realization information when the recognition result shows realization of a sentence. A start/end terminal detection part 1 decides the point of time when the sentence realization information is inputted from the language analysis processing part 4 as the end terminal of the voice when a power duration time exceeds a vocalization time threshold value after the start terminal of the input voice is detected. A voice synthesis part 2 ends the recognition processing of the voice until this point of time, and analyzes the recognition result with respect to grammar, semantics, and the interactive context, and synthesizes the response voice to the input voice to output it. In such a manner, while the input voice is interrupted on the way, the interactive is performed gradually while confirming the contents with a speaker.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a response voice generation method for generating a response voice in response to a user's utterance and a voice dialogue system adopting this method.

[0002]

2. Description of the Related Art Conventionally, there is known a system for automatically recognizing a user's utterance and generating and outputting a response voice corresponding to the utterance content. In such a system, upon detection of the end of the voice uttered by the user, a response voice corresponding to the utterance content is generated, and the generated response voice is output from the voice output device. As a method of automatically detecting the end of the voice in this case, conventionally, when the power of the input voice falls below the minimum power value that can be identified as voiced, if that state is continued for a certain period of time, Was adopted as a voice termination.

Further, the input voice is immediately recognized, and the certainty that the recognition result at the present time is always formed as a sentence is evaluated. Immediately after the certainty that the evaluation value exceeds a preset standard, there is silence. A method has also been proposed in which the state is regarded as the end of input when the state continues for a certain period of time. For this method, refer to the description in "Study on voice segment detection method using likelihood of sentence hypothesis: Naito et al .; 55p" (Autumn Acoustics Society of Japan 1994 Autumn Research Presentation; 2-8-9). .

[0004]

As described above, in the conventional voice dialogue system, when the end of the voice uttered by the user is detected, the response voice corresponding to the utterance content is generated. If it is continuously input for a long time, it takes time to recognize the voice. Moreover, since the result of the voice recognition is understood and the response voice is generated only after the end of the voice is detected, it takes a considerable time until the response to the user's utterance is output, and the accuracy is also lowered. There was a problem that it ended up. At this time, if the system fails to understand the user's utterance content, the user has to utter a long voice again,
There was also a problem that it hindered the progress of the dialogue.

In view of the above problems, an object of the present invention is to improve the understanding rate of a voice in a voice dialogue system which automatically recognizes a voice from a speaker and outputs an appropriate response based on the recognition result. Another object is to provide a technique for reducing system response time and facilitating the progress of dialogue.

[0006]

The present invention provides a response voice generation method that solves the above problems. This method includes at least a step of sequentially recognizing the start of the input voice and recognizing the voice, a step of analyzing whether or not the content of the input voice is satisfied as a sentence based on the result of the recognition, and the start point. Corresponding to the step of deciphering the input voice when the input voice is detected as a sentence at the time when the input voice at the time of detection reaches a predetermined duration and deciphering the content as the end of the input voice. Generating a response voice to perform.

In the step of deciphering the content of the voice, if a silence state is detected before the elapse of the predetermined duration, the content of the voice ending at the point of silence detection may be deciphered. .

The present invention also provides a voice dialogue system suitable for implementing the above-described response voice generation method. This speech dialogue system is an improvement of the conventional method, particularly the method of detecting the end of input speech. Specifically, it is characterized by having the following elements. (1) A first threshold value that defines a minimum power value that can be identified as a voice, a second threshold value that defines a shortest time that can be identified as a continuous voice, and a first threshold that defines a continuous allowable time of continuous voice. Threshold storage means for storing the threshold value of No. 3, (2) The input voice is sequentially recognized upon the input of voice, and it is determined whether or not the sentence is satisfied based on the recognition result. (3) input voice power, means for detecting the power change time of the power, and duration of the voice power before and after the change, and the detected power and voice duration of the input voice. And the first and second threshold values stored in the threshold value storage means are compared to detect the start end of the input voice, and when the voiced duration exceeds the third threshold value or the third threshold value. And received the sentence formation information When, starting and end detecting means for detecting the end of voice at the time of receiving the and the statement holds information beyond the time or the third threshold value exceeds the threshold value of the third,
(4) Decoding the content of the detected sentence from the beginning to the end, detecting a response voice corresponding to the deciphered content, for example, a key phrase of the content of the sentence, and detecting the key phrase Decoding means for synthesizing a response voice including. In addition,
“Voice” refers to a voice state other than the case where the reduction in voice power continues for a certain period of time, and does not necessarily mean only a voice state that exceeds the first threshold value. Also, "Continuous voice"
Is a voice state in which such voiced sound continues in time. Further, it is assumed that the sentence success / failure of the element (2) is determined using, for example, predetermined grammatical information, semantic information, word information, and the like.

The present invention can also use the conventional method together. In this case, a fourth threshold value that defines a silent duration that can be identified as a silent state is stored in the threshold storage unit. Then, the start / end detection means detects power decrease, that is, silence before the sound duration reaches the third threshold, and detects silence when the duration of the silence exceeds the fourth threshold. The time point is detected as the end of the voice.

[0010]

In the start / end detecting means of the present invention, the power of the input voice detected by the voice power detecting means exceeds the first threshold value,
When the voiced duration exceeds the second threshold, the start edge of the voice can be detected. The starting point in this case is a point in time when the voice power exceeds the first threshold value. The voice recognition means is
When the start edge is detected, recognition is started and, as a result of the predetermined analysis and determination processing, sentence establishment information is generated when the sentence is satisfied. The start / end detection means also regards the reception time as the end of the voice when the voiced duration exceeds the third threshold and the sentence formation information is received. When the sentence formation information can be ignored, the time when the third threshold is exceeded is regarded as the end point of the voice. Even before that, when the silence is detected and the duration exceeds the fourth threshold value, the silence detection time is set as the end of the voice. The response voice synthesizing unit decodes the content of the detected sentence from the start end to the end, synthesizes the response voice corresponding to the decoded content, and outputs it. Although the content of the voice uttered immediately after the detection of the end point is not recognized, a gradual dialogue is created by synthesizing a response voice that causes the content of the following unrecognized voice to be input again in the subsequent dialogue. Processing becomes possible.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a configuration diagram of a voice dialogue system according to an embodiment of the present invention. In this voice dialogue system, the voice uttered by the user is input from the voice input terminal IN and guided to the start / end detection unit 1 and the voice synthesis unit 2.

The start / end detection unit 1 includes a voice power detection unit 11 for detecting the power of an input voice and a change in the power, a duration detection unit 12 for detecting (measuring) a voiced duration or a silent duration, and detection values thereof. Various threshold values stored in the threshold table 3, a determination processing unit 13 that performs a predetermined determination process based on the output from the language analysis processing unit 4, and a command signal generation unit 14
Consists of The language analysis processing unit 4 performs language analysis based on the recognition result of the voice recognition unit 21 described later, and outputs the sentence formation information when the recognition result up to the present time can be regarded as a sentence. .

Further, the voice synthesizing unit 2 recognizes a voice inputted from the input unit IN immediately, and a voice recognizing unit 2 for understanding the recognition result and generating semantic expression information.
2. A response generation unit 23 that synthesizes required voices based on the generated semantic expression information to synthesize a response voice,
The synthesized voice is output to the output unit OUT. Input / output control of each of these units is performed by various command signals generated by the command signal generation unit 14.

In this embodiment, the voice recognition section 21 and the language analysis processing section 4 constitute a voice recognition means, and
The speech comprehension unit 22 and the response generation unit 23 are included in the decoding unit.

FIG. 2 is a conceptual diagram showing an example of various threshold values stored in the threshold table 3. In FIG. 2, the voice power threshold 31 is a value (first threshold) that defines the lowest power value that can be identified as a voice, and the voice duration threshold 32.
Is a value (second threshold value) that defines the shortest time that can be identified as continuous voice. These threshold values 31 and 32 are used when the start edge of the input voice is detected. The silent time threshold value 33 is a value (fourth threshold value) that defines the duration time that can be identified as a silent state, and the vocalization time threshold value 34 is a value (third threshold value) that defines the allowable duration of voiced sound. is there. These thresholds 3
3, 34 are used when detecting the end of the input voice. Also,
The maximum vocalization time threshold value 35 is a value that defines the maximum voiced duration that can be accepted by the system.

Next, the operation of each part of the voice dialogue system of this embodiment will be described. FIG. 3 is a procedure diagram of processing centered on the start / end detection unit 1 according to the present embodiment, and S represents a processing step. Referring to FIG. 3, when a voiced state is confirmed by comparing the power value of the input voice with the voice power threshold 31 (S101, 102), the determination processing unit 13 causes the voiced duration to be the voiced time threshold. It is determined whether or not 32 or more continues (S103), and if it continues, the sound detection time point is recognized as the start end of the input voice (S104). Then, the sound duration time is monitored (S105).

Here, the voiced duration is the vocalization time threshold value 3
Before reaching 4 (S105: No), the voice power detection unit 11
Silence is detected, that is, when the sound power starts to fall below the sound power threshold 31 (S106), and the duration detecting unit 12 detects that the silence duration exceeds the silence time threshold 33, the silence detection is performed. The time point is recognized as the end of the voice (S108).

Further, the duration of the voiced time is the vocalization time threshold value 34.
When it exceeds (S105: Yes), the language analysis processing unit 4
It waits until the sentence formation information is input from (S109). When it is input (S109: Yes), or when it can be estimated that the sentence is established even if the non-input at this point is ignored (S109: Yes)
110: Yes), the time when the sentence formation information is received or the time when the utterance time threshold 34 is exceeded is recognized as the end of the voice (S112). In S110, if the non-input of the sentence formation information cannot be ignored (S110: No), the input of the sentence formation information is waited until the maximum vocalization time threshold 35 is exceeded (S111), and if the sentence formation information is input (S111: Yes). ),
The time when the sentence formation information is received is recognized as the end of the voice (S
112).

When the end of the voice is recognized in S108 or S112, the command signal generator 14 is requested to end the voice recognition process and output a command of the recognition result. On the other hand, if the sentence formation information is not input even if the maximum vocalization time threshold value 35 is exceeded in S111 (S111: Ye
s), the command output of the voice input again is performed by the command signal generation unit 14
(S113).

The voice recognition unit 21 terminates the voice recognition process up to the present time in accordance with a command such as the recognition process end sent from the command signal generation unit 14, and passes the recognition result to the voice understanding unit 4. The speech understanding unit 4 analyzes the recognition result grammatically, semantically, and in a dialogue context, and passes the semantic expression information representing the meaning content included in the input speech to the response generation unit 5. The response generation unit 5 generates the semantic expression information of the input voice provided from the voice understanding unit 4 or the synthetic voice corresponding to the voice input command from the command signal generation unit 14 again, and outputs the synthesized voice. Output more. When generating synthetic speech, use the dialogue knowledge and specialized knowledge that are provided in advance, and recognize the main words and phrases that exclude unnecessary words from the contents of the input voice and the response voice that asks the correctness of the words, or the contents of the input voice. Generates audio that could not be done. The above process is repeated until a series of interactive processes is completed (S114).

Next, a specific processing example in the voice dialogue system of this embodiment will be described with reference to FIG. FIG.
(A) is a content explanatory view of the voice uttered by the user, and in the conventional voice start / end detection method, the utterance contents shown in the figure are treated as one sentence, and based on the recognition result, the system A response voice was being generated.

On the other hand, in the present embodiment, (1) through (1)
Speech recognition and content understanding are performed for each section (6), and
A response voice is generated and output in an interactive form as shown in (b). In other words, at the time (1) when the user issues "Well, I want to reserve the conference room ...", the system detects the voiced state and starts recognition. At this time (1), when it is determined that the utterance time threshold value 34 has not been exceeded and the sentence is not yet established, the time "... the meeting room is the third meeting room" is issued.
Continue to recognize until (2). After that, when it is determined that the utterance time threshold 34 is exceeded and the sentence is satisfied, time (2)
It is considered that the utterance for the first sentence is completed, and a response voice corresponding to the meaning contents up to that time, for example, a response voice "third conference room" including the main words and phrases of the input utterance contents is generated and output. . As a result, the user knows that the utterance has been temporarily stopped, and starts uttering the subsequent content, that is, the content after the time (2).

The system side processes the subsequent contents in the same procedure as described above. That is, it is considered that the utterance is completed at the time (4) when the utterance time threshold 34 is exceeded and the sentence is satisfied as the first sentence, and replies, "I hope from 3 o'clock Friday next week." If this response is correct, the user shall report the time (4)
Speak the following contents. If the duration of utterance after the time point (4) (time points (4) to (6)) is less than or equal to the utterance time threshold 34, the system side utters even before reaching the utterance time threshold 34. Is judged to have ended, and when the final vocalization ends
Response voice corresponding to (6) "I want a large conference room."
Is output. In this way, by gradually conducting the dialogue with the user, smooth system operation becomes possible.

[0024]

As is apparent from the above description, according to the present invention, even when a voice is continuously input for a long period of time, the input voice has a predetermined duration (third threshold value).
If it exceeds, or if it is satisfied as a sentence at that time, it is considered as the end of the speech to be sequentially recognized, a required response speech is generated, and the dialogue with the speaker is gradually performed.
In addition to shortening the time until the system response voice is output, the recognition rate of each voice is significantly improved compared to the conventional method. Further, even when the silent state is detected before the above-described duration time elapses, the same processing is performed with the silent point being detected as the end, so that compatibility with the conventional method can be achieved. As a result, the progress of the dialogue with the speaker can be made smooth, and the conventional problems can be solved.

[Brief description of drawings]

FIG. 1 is a block configuration diagram of a voice dialogue system according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of stored contents of a threshold table according to the present embodiment.

FIG. 3 is a procedure explanatory diagram of a start / end detection process of a recognition target voice according to the present embodiment.

FIG. 4 is an explanatory diagram of a concrete dialogue format by the voice dialogue system of the present embodiment, in which (a) is an example of contents of input voice,
(B) is a dialogue procedure diagram based on this content example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 start / end detection unit 11 voice power detection unit 12 duration detection unit 13 determination processing unit 14 command signal generation unit 2 voice synthesis unit 21 voice recognition unit 22 voice understanding unit 23 response generation unit 3 threshold table 4 language analysis processing unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G06F 3/16 330 9172-5E G06F 3/16 330A G10L 5/02 G10L 5/02 J

Claims (5)

[Claims] 1. A step of sequentially recognizing an input voice by detecting a start point of the input voice, a step of analyzing whether or not the content of the input voice is satisfied as a sentence based on the result of the recognition, and the start point detection. Corresponding to the step of deciphering the input voice when the input voice is satisfied as a sentence when the input voice at the time has passed a predetermined duration, and deciphering the content as the end of the input voice. And a step of generating a response voice, the response voice generation method comprising: 2. The step of decoding the content of the input voice, when a silence state is detected before the elapse of the predetermined duration, decodes the content of the voice ending at the point of silence detection. The response voice generation method according to claim 1. 3. A first threshold value that defines a minimum power value that can be identified as a voice, a second threshold value that defines a minimum time that can be identified as a continuous voice, and an allowable duration of a voice. A threshold value storage unit that stores a third threshold value, sequentially recognizes the input voice when triggered by a voice input, determines whether or not the sentence is satisfied based on the recognition result, and when the condition is satisfied, the sentence establishment information is output. The voice recognition means to be generated is compared with the power of the input voice and the duration of the voice and the first and second threshold values stored in the threshold value storage means to detect the start edge of the input voice and detect the start edge. When the voice duration at the time point (hereinafter, voiced duration) exceeds the third threshold value, or exceeds the third threshold value and receives the sentence establishment information, the third threshold value is exceeded. At the time or when the sentence formation information is received Start-end detecting means for detecting the end of the voice, and response voice synthesizing means for decoding the content of the detected sentence from the start end to the end and synthesizing a response voice corresponding to the decoded content. Spoken dialogue system characterized by. 4. A fourth threshold value, which defines a silence duration that can be identified as a silence state, is stored in the threshold value storage means, and the start / end detection means is configured to detect the voiced duration time by the third threshold value. 4. When the silence is detected before reaching, and when the duration of the silence exceeds the fourth threshold value, the silence detection time point is detected as the end of the voice.
The spoken dialogue system described. 5. The response voice synthesizing means detects a key word / phrase of the meaning of the detected sentence from the start end to the terminal end, and synthesizes a response voice including the detected key word / phrase. The voice dialogue system according to claim 3 or 4.