JPH09244680A - Device and method for rhythm control - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a natural speech as a synthesis result when a speech is synthesized by combining a speech of a document different in pronunciation style and a speech of word together. SOLUTION: In addition to a data base 1 wherein speeches of various words are registered and a data base 5 wherein speeches of fixed form parts of various documents are registered, a table 3 containing registered rhythms of those words and a table 7 containing ranges of rhythms that word parts of those documents should have are prepared. The speech and rhythm of a desired word are read out and the speech of a fixed part of a desired document and the rhythm range of the word part are read out. Then the rhythm 23 of the word is compared with the rhythm range 27 of the word part of the document to calculate the difference 29 between both the rhythms. Then rhythm control 17 based upon the difference 29 is exercised over the speech 21 of the word to convert the speech 21 of the word into a speech 31 having a rhythm converging in the rhythm range 27. Then the converted word speech 31 is combined with the speech 25 of the fixed form part of the document to synthesize the speech 33 of the complete document.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention generally relates to speech synthesis technology, and more particularly, it is a natural sentence of a type which is a fixed sentence such as a system response sentence in which only specific words are replaced. The present invention relates to a technique of prosody control suitable for synthesizing various vocal sounds.

[0002]

2. Description of the Related Art A sentence used in a constituent sound of a popular automatic voice response device is generally called a system response sentence, and it is often a fixed sentence in which only a specific word can be replaced. For example, in the system response sentence "There was a transfer from ABC Bank", the word of the bank name was "AB
It can be replaced with a different one such as "C" or "CDF".

When synthesizing such a system response sentence, a word and a sentence are conventionally connected by an edit synthesizing method. The editing and composition method is, for example, the composition of a certain sentence such as "There was a transfer from XX bank" to the word portion "XX" by pasting an arbitrary word such as "ABC" or "CDF". Is the way to do it.

[0004]

[Problems to be Solved by the Invention] In the conventional voice synthesis processing using the edit synthesis method, the voice of a word such as "ABC" and the voice of a fixed part such as "There was a transfer from the bank" in the sentence. And is simply connected. However, even if the same sentence or word is uttered in different utterance styles, the prosody is slightly different, and thus the sensation is auditorily different.
For example, when a voice of a conversation by a plurality of people and a voice of a reading by one person are compared, even the same sentence sounds different because the way of speaking, that is, the utterance style is different.
In addition, even if the sound of a word uttered alone is compared with the sound of the same word uttered incorporated in a sentence, it still sounds different. Usually, the voices of individual words and sentences registered in the voice synthesis system are uttered in different utterance styles. Sounds unnatural.

According to this conventional technique, in order to create response sentences having various utterance styles so as to be heard naturally, other kinds of voices having prosody suitable for various registration sentences are registered for each word. You need to do it. However, especially in a small system, it is not possible to prepare other types of voices for each registered word due to the limited capacity of the database. Also, in terms of voice recording, it is difficult to utter a word with any prosody.

Therefore, an object of the present invention is to make the synthesized result sound natural when the speech of a sentence and the speech of a word having different utterance styles are combined to synthesize a speech.

[0007]

According to the present invention, when a voice of a fixed part of a sentence and a voice of a word are connected to synthesize a voice of a complete sentence, the prosody of the voice of the word is changed to the sentence. Controlled to fit. In this prosody control, first, data indicating the prosody of the voice of the word and data indicating the prosody range of the word portion of the sentence are acquired, and these data are compared. Then, based on the result of the comparison, the voice of the word is adjusted so that the prosody of the voice of the word falls within the prosodic range of the word portion. In this way, the sound of a word whose prosody is controlled is
If you connect it to the sound of a fixed part of a sentence, you get a complete sentence sound that sounds natural.

The prosody parameters operated by the prosody control are:
Three types of parameters can be adopted: pitch frequency, amplitude, and time length. A pitch synchronization waveform superposition method can be used as the prosody control method.

According to the present invention, the voice of a word uttered in one utterance style can be converted into a voice having a prosody suitable for each of various sentences uttered in various utterance styles. There is no need to prepare other kinds of sounds with various prosody. Therefore, even a small system with a small database capacity can synthesize a sentence with a natural voice.

[0010]

1 shows the configuration of a speech synthesizer according to an embodiment of the present invention. This device is actually a computer system in which an application program for speech synthesis processing according to the present invention is installed.

A word database 1, a word prosody table 3, a sentence database 5, and a sentence prosody table 7 are prepared in an appropriate storage such as a magnetic disk device. The word database 1 stores word numbers of various words and voice data obtained from actual utterance of those words. In the word prosody table 3, the prosody data measured from the actual utterances of the various words are
It is stored with the word number. Here, the prosody data of each word is composed of an average value of three types of prosody parameters, which are pitch frequency, amplitude, and time length, which are measured in the voice section of each word.

In the sentence database 5, the sentence numbers of various sentences and the voices of the fixed parts of each sentence (that is, the remaining parts excluding the replaceable word parts) obtained from the actual utterances of those sentences are recorded. Data and are stored. The in-sentence prosody table 7 stores data indicating the range of prosody of the replaceable word parts of the various sentences. The prosodic range data of each sentence is composed of the average value and the standard deviation of the above-mentioned three kinds of prosodic parameters of the word portion of each sentence, and is created as follows. That is, for each sentence, a large number of sentences in which only the word portions are replaced with various words are actually uttered, and the prosody of these word portions (the values of the above three types of prosody parameters) are measured, and Obtain the average value and standard deviation of the measured values for the parameters. According to an experiment conducted by the inventor, it was found that there is a range peculiar to the pitch frequency, the amplitude and the time length of the word part for each individual sentence. Therefore, the prosody range data of each sentence created as described above indicates the prosody range peculiar to each sentence included in the word portion of each sentence. This means that when uttering each sentence, if a word part is uttered in the prosody of the prosodic range data, that word will be auditorily matched with the fixed part of each sentence and the whole sentence will sound natural. means.

By executing the application program, the CPU 9 carries out five processes including a word search process 11, a sentence search process 13, a prosody generation process 15, a prosody control process 17, and an edit synthesis process 19. CPU
Inputs to 9 are word numbers and sentence numbers of words and sentences that form a desired system response sentence.

Word search processing 11 is performed in response to the input word number. In this process 11, the voice data 21 and the prosody data 23 of the word identified by the input word number are searched from the word database 1 and the word prosody table 3. The voice data 21 of the retrieved word is passed to the prosody control process 17, and the prosody data 23 is passed to the prosody generation process 15.

A sentence search process 13 is performed in response to the inputted sentence number. In this process 13, the sentence database 5 and the in-sentence prosody table 7 are searched for the voice data 25 of the fixed part of the sentence and the prosody range data 27 of the word part identified by the input sentence number. The retrieved fixed-form voice data 25 is passed to the edit synthesis processing 19, and the prosodic range data 27 of the word portion is prosody generation processing 15.
Passed to

In the prosody generation processing 15, the prosody data 23 of the word is compared with the prosody range data 27 of the word portion of the sentence, and the difference 29 between them is calculated. For example, the pitch frequency indicated by the prosody data 23 is 277 [Hz], and the average pitch frequency indicated by the prosody range data 27 is 325.
When the frequency is [Hz] and the standard deviation is 34 [Hz], the pitch frequency range indicated by the prosody range data 27 is 291-3.
Since it is 59 [Hz], the difference 29 regarding the frequency is 291-277 = 14 [Hz]. The difference 29 is calculated for other prosody parameters in the same manner.
The difference 29 between these three types of prosody parameters is passed to the prosody control process 17. Note that any of the three types of parameter values indicated by the word prosody data 23 is the prosody range data 27.
, The difference 29 for that parameter is zero.

In the prosody control processing 17, prosody control based on the prosody difference 29 is performed on the voice data 21 of the word. As the prosody control method, for example, the pitch synchronization waveform superposition method is used. As a result of the prosody control processing 17, the original voice data 21 is converted into voice data 31 having a prosody within the prosody range indicated by the prosody range data 27. The difference is 2
For the parameter in which 9 is zero, the original value of the audio data 21 is maintained as it is because no control is performed.
The voice data 31 having the controlled prosody obtained by this processing 17 is passed to the edit synthesis processing 19.

In the edit / synthesis processing 19, the voice data 31 of the word having the controlled prosody is incorporated into the blank word portion of the voice data 25 of the fixed part of the sentence, and the voice data 33 of the complete system response sentence is obtained. Is generated. The voice data 33 is reproduced as voice by a voice output device such as a speaker.

Incidentally, the pitch synchronization waveform superimposing method used in prosody control is characterized in that the synthesized speech as a control result has a high quality and the pitch waveform unit is easily processed. Details of this method can be found, for example, in “Pitch-Syncronous Waveform Processing Techniq” by E. Moulines and F. Charpentier.
ues for Text-to-Speech Synthesis using Diphones, ”
Speech Communication, Vol.9, pp.453-467, Dec. 199
It is described in 0.

FIG. 2 shows the flow of prosody control by the pitch synchronization waveform superposition method, and FIG. 3 shows the speech waveform at each stage of this prosody control.

First, a window function is applied to the original voice data 21 representing the original voice waveform as shown in FIG. 3A to extract individual pitch waveforms as shown in FIG. 3B (S1). Next, a weighting function determined by the difference in amplitude is applied to each pitch waveform to adjust the amplitude of each pitch waveform as shown in FIG. 3C (S2). Next, by adding or subtracting the number of pitch waveforms existing in the voice section according to the difference in the time length, as shown in FIG.
The time length is adjusted as shown in (S3). Next, the pitch frequency is adjusted by changing the pitch waveform interval (cycle) according to the difference in pitch frequency, and by combining these pitch waveforms, a controlled prosody as shown in FIG. 3E is obtained. The voice data 31 representing the voice waveform is stored is created (S4).

By such prosody control, the prosody of the original voice of the word is corrected to be within the prosody range suitable for the sentence to be incorporated, so that the voice of the entire sentence incorporating the word voice naturally. hear. Therefore, one voice data per word can be used by adapting to various sentences having different utterance styles.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention.

FIG. 2 is a flowchart showing the flow of prosody control by the pitch synchronization waveform superposition method.

FIG. 3 is a waveform diagram showing a speech waveform at each step of prosody control.

[Explanation of symbols]

 1 word database 3 word prosody table 5 sentence database 7 sentence prosodic table 9 CPU 11 word search process 13 sentence search process 15 prosody generation process 17 prosody control process 19 edit synthesis process 21 word speech data 23 word prosodic data 25 sentence pattern Partial voice data 27 Prosodic range data of word part of sentence 29 Prosody difference 31 Speech data of words with controlled prosody 33 Synthesized system response sentence

Claims (6)

[Claims] 1. A system for synthesizing a voice of a fixed part of a sentence having a word part and a fixed part and a voice of a word to be included in the word part to synthesize the whole voice of the sentence. A device for controlling the prosody of the word so as to match the sentence, a word acquiring unit for acquiring a voice and a prosody of the word, a prosody range acquiring unit for acquiring a prosody range of a word portion of the sentence, Prosody comparison means for comparing the prosody of the word and the prosody range of the word portion, for the voice of the acquired word, performs prosody control based on the comparison result from the prosody comparison means, thereby, A prosody control device comprising: a prosody control unit that converts the voice of the word into a voice of a controlled word having a prosody existing in the prosody range. 2. The device according to claim 1, wherein the prosody of the word includes three prosody parameter values of a pitch frequency, an amplitude, and a time length of the voice of the word, and the prosody range of the word portion is , The word part has the 3
A prosody control device comprising a range of values of seed prosody parameters. 3. The prosody control device according to claim 1, wherein the prosody control means performs prosody control using a pitch synchronization waveform superposition method. 4. The apparatus according to claim 1, wherein the word acquisition means stores a word database storing sounds of a plurality of words, a word prosody table storing prosody of the plurality of words, and a selected word And a means for retrieving speech and prosody from the word database and a word prosody table, the prosody range acquisition means, a sentence prosody table storing the prosody range of the word portion of the plurality of sentences, and the word of the selected sentence A prosody control device comprising means for retrieving a prosody range of a portion from the in-sentence prosody table. 5. A system for synthesizing a voice of the fixed part of a sentence having a word part and a fixed part and a voice of a word to be included in the word part to synthesize the whole voice of the sentence. , A method of controlling the prosody of the word to match the sentence, a step of acquiring a voice and a prosody of the word, a step of acquiring a prosody range of a word portion of the sentence, of the acquired word The process of comparing the prosody and the prosody range of the word portion, and the prosody control based on the comparison result from the comparison process is performed on the acquired voice of the word, whereby the voice of the word is converted into the prosody. A prosody control method comprising: a process for converting a controlled word having a prosody existing in a range into a voice. 6. A system for synthesizing the entire voice of the sentence by combining the voice of the fixed part of a sentence having a word part and a fixed part with the voice of a word to be included in the word part, A word acquisition unit that acquires a voice and a prosody of a word, a sentence acquisition unit that acquires a voice of a fixed part of the sentence and a prosody range of the word part, and a prosody of the acquired word and a prosody range of the word part And a prosody comparison means for comparing the obtained speech of the word, prosody control based on the comparison result from the prosody comparison means, thereby, the speech of the word, the prosody existing in the prosody range. A prosody control means for converting into a voice of a controlled word, and a voice of the controlled word from the prosody control means;
An edit / synthesis unit that combines the voice of the fixed part from the sentence acquisition unit to create the entire voice of the sentence,
A voice synthesis system characterized by comprising.