JPS63246800A - Voice information generator - 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 [Field of Industrial Application] The present invention relates to a voice information generating device suitable for use in cases where information is exchanged smoothly between humans and machines.

[Summary of the invention]

In this invention, the speed at which speech is generated from a speech synthesis circuit supplied with speech information from a speech information source is selected to be 120±5 m5ec in average mora time, and the speech information for humans is approximately 120±5 m5ec in average mora time. By transmitting at a speed of 5 m5ec, it is possible to speak at a speed that is easy for humans to understand.

[Conventional technology]

In dialogue between humans, the speed of voice information exchange is adapted to the situation where information is exchanged, and this ability to adapt voice speed is thought to play an important role in smooth information exchange. Therefore, if this mechanism for adapting the speed of voice information exchange is introduced to the human-machine interface, it will facilitate the smooth exchange of information between humans and machines, and help realize a human-human interface that respects humanity. It is expected that

Previous research on human characteristics in human-machine voice interaction includes statistical analysis of voice on-off patterns, the nature of the duration of various segments in continuous speech, the critical speed of discrimination in the auditory system, and voice integration. Research includes applications of dynamic synchronization phenomena to humans, machines, and interfaces, and the timing of audio output from machines.

However, in human-machine communication, there is almost no research on the voice response speed from the machine side that is easy to understand for humans who are both speakers and listeners.

[Problem that the invention seeks to solve]

By the way, in conventional speech information generation devices, the speech synthesis circuit only produces sounds at a fixed speed, but the speaking speed is not necessarily easy for the human listener to hear. Also,
The pitch of the vocalizations was also constant, making it difficult to hear.

This invention was made in view of this point, and is a basic research for the development of a system that responds to the human voice input speed at a voice output speed that is easy for the human to understand. By analyzing and evaluating voice information when speaking at a speed that is considered easy to understand, and by examining the individual dependence of voice speed that is easy to understand, we will develop a voice information generation device that can speak at a speed that is easy for humans to understand. This is what we provide.

[Means for solving problems]

The audio information generating device according to the present invention includes an audio information source (5)
, a speech synthesis circuit (7) to which speech information from this speech information source (5) is supplied, and a speed at which speech is generated from this speech synthesis circuit (7) in an average mora time of 120±5.
It is equipped with a speed setting means (9, 10) for selecting m5ec, and the voice information for humans is approximately 12 in average mora time.
It is configured to transmit at a speed of 0±5 m5ec.

[Effect]

Speech information is supplied from an audio information source (5) to a speech synthesis circuit (7) to generate speech. At that time, the speed setting means (9
, 10), the speed at which voice is generated is selected to be 120±5 a+sec in average mora time, and the voice information for humans is approximately 120±5 vasec in average mora time.
transmit at a speed of This makes it possible to speak at a speed that is easy for humans to understand.

〔Example〕

Hereinafter, one embodiment of the present invention will be described using a CD-ROM as an audio information source.
Taking as an example a case where a sound encyclopedia using a ROM is used, a detailed explanation will be given based on FIGS. 1 to 9.

FIG. 1 shows the circuit configuration of this embodiment. In the figure, (1) is a central processing unit (hereinafter referred to as CPU), and a keyboard (2) is connected to this CPU (1).
, ROM (3) and RAM +4).

(5) is a CD-ROM as an audio information source;
This CD-ROM (5) contains explanatory text for each item.
ζ, control signals of different frequencies are recorded together in the normal speed speech part and in parts that are particularly desired to be understood, such as country names, place names, 9 people's names, numbers and 1 hour. This CD-RO
The contents of M(5) are input to the CPU (
1).

Reference numeral (6) denotes a reproducing circuit, which is supplied with information read from the CD-ROM (5), decodes it using an error correction circuit (not shown), and extracts the above-mentioned control signal. (7) is a voice synthesis circuit (7), and the information decoded by the reproduction circuit (6) is taken into a memory (not shown) in the voice synthesis circuit (7), and then output as voice to a speaker. (8) A sound is emitted.

(9) is a control signal detection circuit that detects a control signal from the output of the reproduction circuit (6), and the control signal detected here is supplied to the speed determination circuit (10). Speed determining circuit (
10) controls the 'ζ speech synthesis circuit (9) based on the supplied control signal and determines the speed at which the speech is generated.

For example, if the speech part is at normal speed, the average mora time is 120m5ec, and if it is a part that you want to make clear, such as the name of a country, one place, one person's name, one hour of numbers, etc., the average mora time is 125m.
I will decide to make it sec. Note that this time can be adjusted within a predetermined range, for example, within a range of ±5 m5ec using a volume control (11).

Note that this average mora time does not need to be constant over the entire audio output section, and may be controlled to be shorter at the end of a word, for example.

(12) is a pinch changing circuit, and its output controls the speech synthesis circuit (7) and the speed determining circuit (10). The pitch of the sound can be changed by controlling the voice synthesis circuit (7) with the pitch change circuit (12). Also,
Experience has shown that voices with a short pinch period, that is, voices with a high pitch, sound fast, so the pitch changing circuit (12) controls the speed determining circuit (10) to slow down the voice output speed.

Incidentally, the pinch period in the pitch changing circuit (12) can be adjusted with a volume control (13).

Now, when you input the item you want to hear using the keyboard (2), the CPU (11) controls the CD-ROM (5) and reads the signal from a predetermined disk position.
It is supplied to the ζ regeneration circuit (6). In the reproduction circuit (6), C
The output of the D-ROM (5) is decoded using an error correction circuit and a control signal is extracted.

The signal corresponding to the decoded voice is sent to the voice synthesis circuit (7
) is supplied to the memory thereof, and the control signal is detected by the control signal detection circuit (9).

The detected control signal is supplied to a speed determining circuit (10). The speed determining circuit (lO) determines the speed of the voice generated by the voice synthesizing circuit (7) according to the content of the supplied control signal. In other words, for speech sections at normal speed, the average mora time is set to 120 ± 5 asec, and for parts that are particularly easy to understand, the average mora time is set to 125 ± 5 asec.
Set to m5ec.

In addition, the pinch changing circuit (12) controls the speed determining circuit (lO) and the voice synthesizing circuit (7) as necessary to change the voice output speed and the pinch.

In this way, in this embodiment, it is possible to obtain a device that speaks at a speed that is easy for humans to understand.

In addition, comprehension can be improved by making the generation speed of parts that are particularly desired to be understood, such as country names, place names, names of people, numbers 9 hours, etc., slower than other parts. You can also choose a pitch that you think is easy to pause. Also, you can speak at a speed that matches the speed of your own speech. Furthermore, when the pitch is high, the speech rate can be slowed down to make it easier to understand.

In the above-mentioned embodiment, an analysis circuit (14) is provided to detect a pinch in the human voice produced by the microphone, and its output controls the pitch change circuit (12) to change the output speed and pitch of the voice. In addition, a voice recognition circuit (15) for receiving a microphone signal and performing voice recognition is provided,
The output may be supplied to the CPU (1) and used in place of the keyboard (2).

Next, to ensure smooth information exchange between humans and machines,
The speed of the sound emitted from the machine is set within a predetermined range, that is, in this example, the average mora time is approximately 120±5 m5ec.
I have conducted an experiment that provides the basis for setting this, and I will explain it below.

Figure 2 shows the experimental system diagram. In the figure, (20) is the subject, (21) is the microphone, and (22) is the VTR.
.

(23) is a monitor, (24) is a video timer,
(25) is a video camera, and (26) is a manuscript with utterances written on it. The subjects (20) were 27 males between the ages of 20 and 22. The voice of the subject (20) is captured by the microphone (21).
) to the VTR (22), and at the same time the subject (
20)'s speech is recorded on the VTR (22). The subject (20) himself understood the sentence ``Human-machine systems require the unique characteristics of humans to be incorporated into the machine in terms of workability and safety'' (hereinafter referred to as the ``spoken sentence''). The subjects (hereinafter referred to as ``short-term practice'') were asked to practice speaking at a speed that they thought was easy for them to speak for about 15 minutes.
As a countermeasure for the anxiety of forgetting the memorized utterances, the manuscript (26) with the utterances written on it was displayed below the video camera (25), and the participants (20) ) spoke to the video camera (25) ``OO University Faculty of Engineering, Department of Information Technology (hereinafter referred to as ``Voice IDJ'') student number, name,'' and after a few seconds, the subject (20) himself did not understand the sentences he had practiced for a short period of time. I thought and spoke quickly. The subject (20) was checked to see if he was able to speak at a speed that was easy to understand, and the recording was repeated up to three times until a certain level of consent was obtained.

Next, from among the 27 subjects, we selected 10 subjects (see Figure 5) who had a large difference in the uttered sentence interval Tse (see Figure 4) from each other based on the results of analysis using the analysis method described later. After the long-term proficiency practice, the subjects were asked to practice for about 90 minutes until they were completely proficient at the speaking speed they thought was easy to pronounce. Similarly, "Voice ID, Academic Number, Name, Vocalized Sentence" was generated, and the recording was repeated until the subject was completely satisfied.

FIG. 3 is a diagram showing the configuration of the analysis system. In the same figure,
(30) is a microcomputer, (31) is a hard disk, (32) is a display, (33)
) is an A/D converter, and (34) is a clock generator.

(35) is a low-pass filter, (36) is an audio amplifier, (37) is a VTR, and (38) is a monitor.

This system consists of a combinator (30), a low-pass filter (35), and an A/D converter (3) with a resolution of 12 bits.
3) is added, and the sampling frequency is approximately 1 at a sampling frequency of 24kHz.
This system can record and analyze one second of acoustic data.

After A/D converting the acoustic data, it is possible to extract all or a necessary part and analyze the pitch period, etc.

The audio information recorded on the VTR (37) is recorded in this system for each subject. The interval between the audio output section of the audio signal was automatically detected based on the duration distribution of the audio signal below a certain threshold, and the detection result and the audio waveform were measured by visually checking. An example of the audio signal analysis is shown in FIG. 4. The section where the square wave is high is the audio output section, and the section where the square wave is zero is the section in between. The mouth part of the uttered sentence "Human-machine systems are verbal..." is the part where one comma is added to the utterance, so this part was selected as a representative interval Tpa (hereinafter referred to as Tpa). , simply written as ``ma''). On the other hand, the "characteristic unique to humans" of uttering a voice in one breath without a short pause was selected as a representative voice output section Tph (hereinafter simply referred to as the "voice output section"). In addition, the "a" of "workability" in the uttered sentence was calculated using the cepstrum as a representative pinch period.

The relationship between the uttered sentence section Tse and the uttered ID section Tid is expressed as the fifth
In the 0 diagram shown in the figure, the * mark represents short-term training (27 people), and the Δ mark represents long-term mastery training (10 people). Both axes are standardized by the mean value and 3 standard deviations of the analyzed data during short-term practice.

The correlation between the two in short-term practice is 0.67. To test the hypothesis of population correlation coefficient p=o, if we use a two-tailed test on the distribution of It is rejected. Since there is a significant correlation between the utterance ID section that is unconsciously uttered and the utterance sentence section that is consciously uttered, it is considered that humans unconsciously utter at a speed that is easy to understand. The average mora time of the 27 subjects in this short-term practice (length of time equivalent to one character when speed speech is written in kana, here uttered sentence interval / 62) is:
136±16.8 (standard deviation) msoc. Furthermore, the average mora time of the 10 subjects (indicated by Δ) before long-term familiarization practice was 140±18.5 s+sec. On the other hand, in the uttered sentence section of long-term mastery practice, °ζ is the same as for one subject D (
With the exception of 3 standard deviations of short-term practice, and the subject self-judged himself to be a little slow, the results were within a certain time range. This average mora time is 158±10■se
It is c.

The average mora time (i.e., not including pauses) of the 27 subjects during short-term practice in the audio output section Tph was 118±
Subject 10 before long-term familiarization exercise at 11.7m5ec
The average mora time for humans is 115±9.1 m5ec.

On the other hand, the average mora time excluding subjects during long-term proficiency practice was 120 ± 5.0 (standard deviation) @sec, and as with the average mora time of uttered sentences, the average mora time fell within a certain range as speech proficiency was acquired. ing.

From this, it is thought that there is a speech rate common to humans that is easy to understand, and this average mora time can be used as a guideline for the speech synthesis output speed that is generally easy to understand.

The relationship between the voice output interval Tph and the interval Tpa is shown in FIG. 6. The correlation coefficient between the two in short-term practice is extremely low at -0.02 (there is no correlation between the voice output speed and the interval Tpa).

Furthermore, there is no correlation between the two in long-term familiarization practice. Therefore, there is no significant correlation between the voice output speed and the speed, depending on the individual.

During short-term practice, the pitch period of “a” and “
The correlation coefficient with the pitch period of "A" was 0.91.Here, the pinch period of "A" was taken as a representative pitch period.The relationship between the audio output section Tph and the pitch period Tpi was As shown in the figure, the correlation coefficient between the two in short-term practice is -0.4
1, and there is a significant negative correlation between the audio output section and the pinch period. That is, there is a positive correlation between the audio output speed and the pinch period. This seems to contradict our experience in everyday life, but under the condition of an easy-to-understand speech rate,
Experience has shown that sounds with short pinch cycles sound faster, which is interpreted as feedback that slows down the sound output speed. Furthermore, although the voice output speed has slowed down due to long-term proficiency practice, there is no constant tendency in the fluctuation of the pinch cycle.

The relationship between pitch Tpa and pitch period Tpi is shown in FIG.
The correlation coefficient between the two in short-term practice is as low as -0.09, and there is no correlation. Even in long-term proficiency practice, the correlation coefficient is -0.06, excluding the above-mentioned subject, and there is no correlation.

For short-term practice (27 people) and long-term mastery practice (10 people), the correlation coefficients of the utterance ID interval, utterance sentence interval, pause, voice output interval, and pitch period are shown in Figure 9. The correlation coefficients shown in the lower part of the exercise are for the case where subjects who self-judged that the utterance intervals in the uttered sentences were particularly long and somewhat slow were excluded.

In short-term practice, the correlation coefficient between the uttered sentence section and the pitch period is -0.33, which is a lower degree of negative correlation than the correlation coefficient (-0.41) between the speech output section and the pitch period.

This is because the uttered sentence includes a pause that has no correlation with the pinch cycle. Further, the correlation coefficient between the utterance ID section and the pitch period is -0.50, which is a significant correlation. This is interpreted to be because, as in the case of Figure 7, the person unconsciously utters the voice at a speed that is easy to understand.

In the lower correlation coefficient (N=9) for long-term mastery practice,
The threshold value of r at a significance level of 5% for degrees of freedom N-2 = 7 is 0.
．． 666, the hypothesis of the correlation coefficient p=o is not entirely rejected. This is because the utterance interval of the uttered sentence has fallen within a certain range through long-term familiarization practice.

In this way, in this experiment, we analyzed and evaluated the easy-to-understand speech rate of people who are both speakers and listeners, and examined the individual dependence of easy-to-understand speech speed as people become more proficient in speech. As a result of the analysis, the following conclusions were reached, provided that the speaker speaks at a speed that he or she considers easy to understand.

l) There is no correlation between audio output speed and speed.

2) There is a positive correlation between audio output speed and pitch period.

3) The rate of speech that is easy to understand depends on the individual, but once a person becomes familiar with the speech, the rate of speech falls within a certain range that is common to humans and is easy to understand (principle of equal length of mora due to learning).

〔Effect of the invention〕

As described above, according to the present invention, the speed at which speech is generated from the speech synthesis circuit supplied with speech information from the speech information source and the average mora time are selected to be 120±5 tasec, and the speech information for humans is selected. Since the information is transmitted at a speed of approximately 120±5 a+sec in average mora time, it is possible to obtain an audio information generating device that speaks at a speed that is easy for humans to understand. Furthermore, comprehension can be improved by making the generation speed of parts that are particularly desired to be understood, such as country names, place names, 9 people's names, and 9 hour numbers, slower than other parts. In addition, you can choose the speed of your speech that you think is difficult to open, you can adjust the speed of your speech to match the speed of your speech, and you can also slow down the speed of your speech when the pitch is high. It can be made easier to understand.

[Brief explanation of the drawing]

FIG. 1 is a circuit diagram showing an embodiment of the present invention, and FIGS. 2 to 9 are diagrams for explaining the basic principle of the invention. (1) is the central processing unit, (5) is the CD-ROM, (6)
(7) is a speech synthesis circuit, (9) is a control signal detection circuit, (10) is a speed determination circuit, and (12) is a pitch change circuit.

Claims (1)

[Claims] A voice information source, a voice synthesis circuit to which voice information from the voice information source is supplied, and a speed at which the speed at which voice is generated from the voice synthesis circuit is selected to be 120±5 msec in average mora time. 1. A voice information generating device, comprising a setting means, and configured to transmit voice information to a human at a speed of approximately 120±5 msec in average mora time.