JPH0731509B2 - Voice analyzer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice analysis system in which voice is decomposed into phrase units and stored, and particularly, a natural and large amount of reproduced voice such as a guide broadcast is required. It relates to a voice analysis method suitable for use in creating a voice pattern for use in the field.

[Conventional technology]

Conventionally, a speech synthesizer as disclosed in Japanese Examined Patent Publication No. 59-22240 is known. When a sentence number is designated, this speech synthesizer reads a series of bunsetsu addresses that compose the voice sentence from the sentence table, reads bunsetsu from the voice pattern memory based on this bunsetsu address, decodes the sentence, and outputs the sentence. And requires a lot of voice patterns. A method disclosed in Japanese Patent Application Laid-Open No. 58-205196 is known as a method for creating this voice pattern.

In this conventional technique, a signal below a preset reference level continues for a preset reference time or more in a silent section of the data uttered for each phrase or the data uttered a sentence with an appropriate interval between phrases. Then, the voiced part is cut out as a voice pattern.

[Problems to be solved by the invention]

However, when the voice synthesizing device is operated by using the voice pattern created by the above-mentioned conventional technique to perform the guide broadcasting, the vocalization rate and the tone are unnatural. That is, the above-mentioned conventional technique relies on the detection of silent intervals for segmentation of bunsetsu, so the original sound is not in the form of a sentence, but must be uttered in bunsetsu units, or must be uttered with an interval longer than necessary, There was a problem that the speech pattern of the bunsetsu that could give a natural reproduced speech in terms of accent, intonation, and interval was not obtained. In addition, when a silent part is generated in a phrase, a wrong phrase is cut out, and a confirmation and correction operation is required.

An object of the present invention is to solve the above problems and provide a voice analysis method capable of obtaining a phrase-corresponding voice pattern that can synthesize a natural reproduced voice when used in a voice synthesizer. To do.

[Means for solving problems]

According to the present invention, the object is to utilize the syntactic information about the bunsetsu to be cut out, and the number of syllables in the utterance part and the utterance part length are calculated in the syntactic information about the bunsetsu from the speech data naturally uttered in sentence units. Of the corresponding phrase is judged whether or not the condition is equivalent, and if the condition is satisfied, the voiced part is detected as a voice pattern corresponding to the corresponding phrase, and if the condition is not satisfied, the condition is satisfied. This can be achieved by merging adjacent syllables up to this point, or separating the last syllable in the voiced part, and searching for a voiced part that satisfies the condition and detecting it as a voice pattern.

[Work]

The syllable extraction according to the present invention is performed by extracting a voiced sound such as a vowel, and the syllable detection is performed by the number of syllables. The number of syllables in the detected syllable is less than or equal to the number of syllables in the actual syllable due to the undetectable unvoiced sound.Unless the original utterance is unclear, the number of syllables in the detected syllable is
No more than the actual number of syllables.

Further, the syllable length of the original sound fluctuates about ± 10% for each utterance, but it does not change much more, and the present invention cuts out the syllable according to this syllable length and the number of syllables described above. You can easily separate the correct phrase.

Furthermore, since it is possible to use speech data that is naturally uttered in sentence units as the original sound, when speech synthesis is performed using the speech patterns of the separated clauses, a very natural reproduced speech can be obtained.

〔Example〕

An embodiment of a voice analysis system according to the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram of a speech analysis apparatus for implementing a speech analysis method according to the present invention, FIGS. 2 to 5 are diagrams showing a memory table for explaining syntax information, and FIG. 6 is a diagram showing an input speech waveform. FIG. 7 shows a line and FIG. 7 is a comparison / determination table of voice data and phrase information. In FIG. 1, 1 is a microphone for inputting an original sound to be analyzed, 2 is a voice encoding device for converting analog data from the microphone 1 into ADPCM etc., and 5 is a control device for analyzing voice by a micro program. , 6 is a control memory that stores a microprogram for operating the control device 5, 7 is a buffer memory that stores the voice data encoded by the voice encoding device 2, and 8 is a voice pattern for each analyzed phrase. Is a voice pattern memory for storing, and 9 is a syntax information memory for storing syntax information regarding syllables and syllables necessary for analysis,
3 is a voice decoding device for decoding voice data from the voice pattern memory 8 into an analog signal, and 4 is a voice decoding device 3
Is a speaker for outputting the output signal of as a voice, and 10 is an operation console for inputting and displaying the syntax information and reporting the analysis progress.

The control device 5 in the voice analysis device shown in FIG. 1 is controlled by a microprogram in the control memory 6, stores the digital voice data of the original voice from the voice encoding device 2 in the buffer memory 7, and stores it in the syntax information memory 9. Based on the syntax information about the phrase, the phrase is separated from the voice data in the buffer memory 7, and the voice analysis operation of storing the phrase in the voice pattern memory 8 is performed. In the syntax information memory 9, a syllable information table is prepared in advance, which shows the type of syllable and the average required utterance time of the syllable.

Now, for example, it is assumed that the user wants to obtain a speech pattern of a clause for synthesizing the sentences of the sentence numbers 1 and 2 shown below by a voice-sounding device.

1. "Soon, a train bound for Tokyo will be on the first line." 2. "A train bound for Osaka will soon be on the second line." First, the operator uses the console 10 to delimit kana characters and punctuation marks. Use "/" to enter the sentence you want to synthesize. Based on the microprogram stored in the control memory 6, the control device 5 stores the sentence input from the console 10 in the syntax information memory 9 as an input sentence table as shown in FIG. Next, the control device 5 decomposes the contents of the input sentence table for each clause, and if there are the same clauses, they are combined into one clause, and the clause data table as shown in FIG. 3 is used as syntax information. It is generated in the memory 9. The control device 5 uses the syllable information table, and for each syllable in the syllable data table, the total number of syllables included in the syllable and the average utterance required time of the syllable,
That is, the average utterance required time of the phrase is calculated, and the phrase information table as shown in FIG. 4 is generated in the syntax information memory 9. Further, the control device 5 generates a sentence information table as shown in FIG. 5 based on the phrase number sequence indicating the sequence of the phrases for forming each sentence, and stores it in the syntax information memory 9.

When the above processing is completed, the control device 5 sequentially displays the sentences of the sentence number 1 on the console 10 and prompts the user to input the voice of the displayed sentence.

Therefore, the operator or a dedicated announcer is the microphone 1
When the relevant sentence is naturally generated and input at a predetermined speed, the input voice signal is encoded by the voice encoding device 2 and stored in the buffer memory 7. At this time, the input voice signal has a waveform as shown in FIG. 6, for example, and simultaneously with the input to the buffer memory 7, the positional information of the sound part 11 and the syllable 12 is also stored together with the input voice signal.

Each time a sentence is input by voice, the control device 5 sequentially extracts the voiced part stored in the buffer memory 7 from the first voiced part, and outputs the voiced part from the syllable information table in the syntax information memory 9. And the phrase information table, and it is determined whether the sound part is equivalent to a phrase. For example, in the case shown in FIG. 6, first, the sound part 11 corresponding to "mamona"
Are extracted from the buffer memory 7, and the number of syllables and the utterance time included in the voiced part are compared with the number of syllables of the phrase number 1 shown in FIG. 4 and the required utterance time. This comparison is the seventh
The first comparison determination table (a) and the comparison determination tables (b) of the second and subsequent stages shown in the figure are used. In the tables (a) and (b),-is a buffer memory 7
Separate the last syllable element from the voiced part extracted from, + adds the next voiced part to the voiced part extracted from the buffer memory 7, ○ indicates the voiced part extracted from the buffer memory 7 is equivalent to a phrase Data (hereinafter referred to as phrase)
Δ is the same as ◯, but if the judgment in the previous stage is a position with *, the judgment is that the sound part in the previous stage is regarded as a phrase.

In the above case, "Mamona" retrieved from the buffer memory 7
Since the number of syllables and the utterance time of the vowel part are smaller than the number of syllables and the required utterance time of the syllables in the verses information table, the control device 5 controls the + in the lower right corner of the table (a) of FIG. That is, the next voiced part is added to the voiced part extracted from the buffer memory 7, the next voiced part “K” is extracted from the buffer memory 7, and “Mamona”
7 is added to the voiced part, and comparison judgment is made based on the table (b) of FIG. As a result, the sound part “Mamona Ku” extracted from the buffer memory 7 is close to the phrase with the phrase number 1 in the phrase information table of FIG. It is transferred to and stored in the voice pattern memory 8. After that, the control device 5
In the same manner, the phrase corresponding to the phrase is cut out from the buffer memory 7 and stored in the voice pattern memory 8 one after another. Finally, the contents of the voice pattern memory 8 are decoded by the voice decoding device 3, output as synthesized voice from the speaker 4, and confirmed by the operator.

In the above-described embodiment, even as it is, the sentence information table shown in FIG. 5 is searched by the output request information by the sentence number from the console 10 or the higher-level device to replace it, and the sentence for which the voice output request is made is constructed. Based on the phrase number of the phrase, the voice phrase stored in the voice pattern memory 8 can be read, decoded by the voice decoding device 3, and used as a voice synthesizer capable of being output from the speaker 4, and used as a voice analysis and synthesizer. it can.

Furthermore, if the automatically generated sentence information table and the voice pattern corresponding to the phrase of the analysis result are transferred to another voice synthesizer, the voice synthesizer can immediately operate,
Conventionally, in a speech synthesizer, it becomes unnecessary to create a table corresponding to a sentence information table in addition to the analyzed speech pattern data.

〔The invention's effect〕

As described above, according to the present invention, by using the original sound naturally uttered in sentence units, based on the syntactic information, it is possible to automatically separate the speech patterns corresponding to the clauses and perform the speech analysis. When this voice pattern is used in a voice synthesizer, a very natural reproduced voice can be obtained. In addition, since data that is naturally uttered in sentence units is used as the original sound, it is possible to reduce the burden on the utterer of the original sound.

[Brief description of drawings]

FIG. 1 is a block diagram of a speech analysis apparatus for implementing a speech analysis method according to the present invention, FIGS. 2 to 5 are diagrams showing a memory table for explaining syntax information, and FIG. 6 is a diagram showing an input speech waveform. FIG. 7 is a diagram showing an example, and FIG. 7 is a diagram showing a comparison determination table. 1 ... Microphone, 2 ... Voice coding device, 3 ... Voice decoding device, 4 ... Speaker, 5 ... Control device, 6 ... Control memory, 7 ... Buffer memory, 8 ... Voice pattern memory, 9 ... Syntax information memory, 10 ... Operator console

Claims (1)

[Claims] 1. A control device for analyzing and controlling a voice uttered in sentence units according to a program stored in a control memory, a voice encoding device for converting voice input information into digital information, and the voice encoding device. A buffer memory for storing digital information output from the device, a voice pattern memory for storing the voice analyzed by the control unit as digital information in phrase units, and a phrase of an input sentence,
An average utterance required time for each phrase, a syntax information memory for storing the number of syllables in a phrase in advance, in a voice analysis device for creating voice synthesis information in units of phrases, the control device,
A method for decomposing a sentence to be composed, which is input from the input device of the operator console, for each clause, and combining the same clauses into one clause, and for each clause obtained by the decomposition Along with assigning the phrase number, the average utterance required time of each phrase and the number of syllables in the phrase are assigned, and the input sentence to be synthesized, means for creating the phrase number by editing the phrase number, When a sentence created by the phrase number is stored in the syntax information memory and an actual voice of a sentence to be synthesized is input, the input voice is converted into digital information by the voice encoding device. Means for storing in the buffer memory, and stored in the buffer memory so as to satisfy the conditions of the average utterance required time of the syllable stored in the syntax information memory and the number of syllables in the syllable. And a means for detecting the voiced part from the digital information of the input voice and dividing the digital information of the input voice into phrase units, and a means for storing the divided digital information of the voice of the phrase unit in the pattern memory. A voice analysis device characterized by the following.