JP2006133559A - Combined use sound synthesizer for sound recording and editing/text sound synthesis, program thereof, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To put together the recording sound and the sound synthesized from a text to high quality. <P>SOLUTION: This sound synthesizer is built, by using a recording sound database which stores voices for the same contents but in different talking tones, a text-analyzing section to identifies the recording sound matching text which can use the recording sound stored in the database from the inputted texts, a recording sound search section for deciding the tone components over the whole target text by a statistical method and to choose the recording sound, on the basis matching the tone components and the phoneme states, before and after and that for choosing the recording sound continuing, as long as possible with priority, a rhythm generator for generating rhythms from consistency with the tone components, the fundamental frequency, the speed, and the power of the selected recording sound in the texts other than the identified texts, and a sound element search section to generate the sound, by synthesizing the text sound by using the generated rhythm. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  In speech synthesis, when creating a voice by combining a standard part using recorded voice and a variable part corresponding to the synthesized voice, a plurality of recorded voices used in the standard part are prepared and the most natural The present invention relates to a recording / editing / text-to-speech combined speech synthesizer that selects an appropriate recorded speech to obtain synthesized speech, a program, and a recording medium that records the program.

One of speech synthesis technologies for generating desired speech by inputting a sentence (text) to be uttered by a machine or the like is recording / editing speech synthesis. This speech synthesis method is currently used for the voice guidance that flows at the station platform, and is achieved by connecting the recorded speech in units of semantic lengths such as phrases.
In this method, noise may occur due to acoustic elements such as a gap in the fundamental frequency at a connection portion where recorded voices are connected. Documents describing techniques for avoiding this include (Patent Document 1) and (Patent Document 2).

Since the methods disclosed in these documents are basically recorded voices, high quality voices can be created. However, since the contents to be uttered must be recorded again, the maintenance cost is reduced. There is a problem that it is expensive.
On the other hand, there are various other speech synthesis methods such as waveform connection type speech synthesis (Patent Document 3) or parameter speech synthesis, and the quality of synthesized speech has become practical enough to withstand practical use. , Not as much as the recorded live voice.
As other speech synthesis methods, a recording / editing speech synthesis and a waveform-connected speech synthesis capable of synthesizing arbitrary text are conceivable. This method is effective in scenes where the utterance content is limited to some extent, and creates a synthesized speech by mixing a fixed portion using recorded speech and a variable portion using synthesized speech. In this synthesizing method, if the speech of the standard part is recorded in advance and defined as continuous speech, sentences that can be output as synthesized speech are determined indefinitely, and there is a disadvantage that the versatility is lacking.

One way to avoid this is to define the fixed part as a smaller unit. Then, determine whether there is a part that matches the recorded voice in the text you want to synthesize. If there is a matching part, select the recorded voice. If not, use the synthesized voice as the synthesized voice. Since it can be used flexibly according to the situation, versatility increases.
In this synthesizing method using both recording editing synthesis and speech synthesis, phrases that frequently appear in a task whose utterance content is limited to some extent are recorded in advance. Thus, since the recorded voice is used at a high frequency when synthesizing the voice, it is possible to create a higher quality voice than the synthesized voice created only from the voice synthesis.

The method of outputting the sound by connecting the fixed part and the variable part has a problem that the quality of the synthesized speech seems to be low due to the difference in the sound quality between the fixed part and the variable part and the mismatch in the acoustic characteristics of the connection part. .
In other words, in order to create a higher quality synthesized speech, it is necessary to fill in the difference in quality between the fixed portion composed of the recorded speech and the variable portion composed of the synthesized speech. In order to solve this problem, in the speech synthesis method realized by mixing the standard part using the recorded speech and the variable part using the synthesized speech, the acoustic feature of the synthesized speech is changed to match the acoustic feature of the recorded speech. In addition, attempts to improve the quality of the entire synthesized speech by reducing various mismatches between the recorded speech and the synthesized speech as much as possible have been considered.
JP-A-7-56597 Japanese Patent No. 3035318 Japanese Patent No. 2761552

However, if the combination of recorded voices with a synthetic voice in between is not good, for example, if the pitch, that is, the difference between the fundamental frequencies is too large, it will not be heard naturally. This adversely affects the synthesized speech that is subjected to the combined processing, resulting in the inconvenience of creating an unnatural synthesized speech.
SUMMARY OF THE INVENTION An object of the present invention is to propose a recording / speech / text-to-speech combined speech synthesizer capable of creating a high-quality synthesized speech in a speech synthesis method using both recording / editing synthesis and speech synthesis. .

  The recording / synthesizing / text-to-speech combined speech synthesizer according to the present invention stores a plurality of voices having the same utterance content but different tone as a selectable unit of a voice fragment composed of at least one accent phrase. A recorded voice database, a text analysis unit for identifying a recorded voice-compatible text that can use the recorded voice stored in the recorded voice database with respect to the input text, and the identified recorded voice-compatible text, As a condition for selecting a plurality of recorded voices that can be used, the speech component is extracted from the recorded speech database, and the speech adjustment for the entire target sentence is determined by a statistical method. Select recording sound based on phoneme environment match, and select recording sound with priority as much as possible. A voice search unit, and a prosody generation unit that generates prosody based on consistency with the tone component, fundamental frequency, speech speed, and power of the selected recorded voice in a text other than the identified recorded voice-corresponding text. A speech segment search unit that generates synthesized speech by text speech synthesis from the prosody, and a synthesis processing unit that performs synthesis processing as necessary to generate smooth synthesized speech.

  According to the present invention, when selecting a recording voice in a recording / editing / speech-synthesizing speech synthesizer that uses both recording-editing speech synthesis and text-to-speech synthesis capable of creating speech from arbitrary text, High-quality synthesized speech can be obtained by selecting optimal speech data from multiple candidates in consideration of the consistency of the speech component of the entire sentence and the matching of the front and back phoneme environment with adjacent speech segments. it can.

It is conceivable that the recording / synthesizing / text-to-speech combined speech synthesizer according to the present invention may be configured by hardware, but the recording / synthesizing / text-to-speech combined speech synthesizing program according to the present invention is installed in the computer. The best mode is one in which the voice synthesizer is used as a voice synthesizer combined with recording / editing and text-to-speech synthesis.
In order for a computer to function as a voice synthesizer that uses both recording and editing and text-to-speech synthesis, the computer can select a voice fragment composed of at least one accent phrase, and can produce voices with the same utterance content but different tone. Recording audio database that has been saved multiple times,
A text analysis unit that identifies recorded voice-compatible text that can use recorded voice stored in the recorded voice database with respect to the input text, and recording that can be used for the recorded voice-corresponding text in the identified recorded voice-compatible text The speech component is extracted from the recorded speech database as a condition for selecting speech from a plurality of candidates, the amount of speech adjustment of the target sentence as a whole is determined by a statistical method, and the speech component is matched with the front and back phoneme environment. The recording voice search unit that selects the recorded voice with priority as much as possible, and selects the recorded voice with priority as much as possible, and the tone of the selected recorded voice in the text other than the identified recorded voice compatible text Prosody generation unit that generates prosody from consistency with components, fundamental frequency, speech speed, and power, and text speech synthesis from the generated prosody A speech unit searching unit for generating a formed sound, the synthesis processing on if necessary, to build a synthesis processing unit that generates a smooth synthetic speech. By adopting this configuration, it is possible to obtain the unique effect of the present invention that a high-quality synthesized speech can be obtained.

  First, FIG. 1 shows a basic configuration diagram of the present invention. The basic flow of the present invention is to determine which part of the input text 11 in the inputted input text 11 is prepared as the recorded voice in the recorded voice database 14 and select the output voice from the recorded voice. The correspondence text and the non-corresponding text that creates synthesized speech by text-to-speech synthesis are assigned to each processing flow, and the morphological analysis and reading / accent / tone combination type are added to the non-corresponding text. , Continuity of speech tone components and recorded speech obtained by extracting from the recorded speech database 14 by the speech tone component extracting portion 15 or giving them manually, before and after on the input text The recording voice search unit 13 selects the recording voice for the text corresponding to the recording voice based on the coincidence of the phoneme environment. In order to synthesize text-to-speech text that does not support recorded speech, the speech feature extraction unit 17 extracts the speech tone component, fundamental frequency, speech speed, power, etc. of the recorded speech selected by the recorded speech search unit 13. The prosody generation unit 18 generates a prosody using the generated speech feature, and the speech unit search unit 19 performs a process of selecting a speech unit from the general-purpose speech database 20 based on the generated prosody. 13. The synthesized speech data 22 is created by smoothly connecting the speech selected by the speech segment search unit 19 by the synthesis processing unit 21.

Each processing of the speech tone component extraction unit 15 and the recorded voice search unit 13 which are features of the present invention will be described in detail below.
The tone component extraction unit 15 sets the tone component of the entire sentence of the synthesis target text. First, what kind of speech tone component the voice in the recorded voice database 14 actually has is extracted. Speech tone components are extracted for each accent phrase. An example will be described with reference to FIG. 2 in which the horizontal axis represents time and the vertical axis represents the fundamental frequency.
When the basic frequency F1 at the beginning of the accent phrase and the basic frequency F2 at the end of the accent phrase are connected by a straight line N, and the frequency pattern PT after the middle point of the accent phrase is below the straight line N, the frequency pattern after the middle point passing through the starting point Draw a tangent line J to PT. The slope of this tangent line J is used as the speech tone component.

Next, the speech tone component extracted from the recorded voice database 14 is statistically analyzed, and the speech tone component of the synthesis target text is set. As a first step, a normal distribution of the start frequency and end frequency of the accent tone speech component and the slope of the tone component is created for each same accent phrase. The value and slope of the start frequency of the line segment representing the speech component are set so that the sum of the probabilities of the normal distribution is maximized. As a result, a sample tone component can be created.
If the speech tone component created by such a method is used as a sample, the recorded speech search unit 13 can obtain a large number of appropriate recorded speech selection candidates that are close to the sample speech tone component. It becomes possible to create.

The recorded voice search unit 13 selects an appropriate recorded voice in order to improve the synthesized voice finally obtained. First, it is searched whether or not the recorded voice recorded for as long as possible corresponding to the recorded voice corresponding text exists in the recorded voice database 14. If there is a continuous recording voice including two or more accent phrases, in the present technique in which one accent phrase is basically handled as one fragment voice, it is treated as one fragment voice.
Next, the speech tone component data 16 of the recorded speech selection candidate is extracted by the method used in the speech tone component extraction unit 15, and how much the speech tone component matches with the sample speech tone component is shown in FIG. The probability is obtained from the distribution group, and the average probability per fragment speech is calculated.

Also, the degree of coincidence of the phoneme environment on the input text of the fragment speech is calculated as a score. In this calculation method, for example, scoring is performed such that 1 if phonemes are completely matched, 0.5 if phonemes such as vowels and plosives are matched, and 0 otherwise. Then, the combination of the recorded speech having the highest integrated score represented by the weighted addition of the probability based on the tone component and the score based on the front and back phoneme environment is selected.
This is a speech synthesis method that combines recording and editing speech synthesis, which creates a desired speech by connecting recorded speech data in advance and text speech synthesis, which can create any speech from text. What inconveniences occur when the invention is not applied will be described with reference to FIGS.

When a plurality of recorded voices having the same utterance contents are not held in the database, or when a recorded voice is selected without considering non-adjacent recorded voices, there is a high risk of a large continuous gap as shown in FIG. Further, when prosody transformation is performed on the synthesized speech as shown in FIG. 5 in order to fill the continuous gap, the quality of the synthesized speech deteriorates because the prosody of the synthesized speech becomes unnatural.
The present invention shows an example by considering the consistency of the tone component of the entire sentence, the continuity of the recorded speech, and the matching of the front and back phoneme environment with adjacent speech segments when selecting the recorded speech. It is an effective means that can avoid such quality degradation.

  The recording / synthesizing / text-to-speech combined speech synthesizer according to the present invention described above mainly installs the recording / synthesizing / text-to-speech combined speech synthesis program according to the present invention in a computer, and the recording / editing / text-to-speech combined type to the computer. This is realized by functioning as a speech synthesizer. The recording / speech / text-to-speech combined speech synthesis program according to the present invention is written in a computer-readable program language and recorded on a computer-readable recording medium such as a magnetic disk, a CD-ROM, or a semiconductor memory. From these recording media, it can be installed in a computer or installed through a communication line.

  The recording / editing / text-to-speech combined speech synthesizer according to the present invention has good voice quality, and is used for various automatic voice guidance devices or automatic response devices.

The block diagram for demonstrating one Example of this invention. The graph for demonstrating operation | movement of the Example shown in FIG. The same graph as FIG. The graph for demonstrating the inconvenience of the audio | voice synthetic | combination type | formula voice-synthesizing combined use of recording edit which does not implement this invention. The same graph as FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Input text 18 Prosody generation part 12 Text analysis part 19 Speech segment search part 13 Recorded voice search part 20 General-purpose speech database 15 Speech tone component extraction part 21 Synthesis processing part 16 Speech tone component data 22 Synthetic voice data 17 Speech feature extraction part

Claims (3)

Sound recording / synthetic speech synthesis combined speech recording combined with voice recording that creates desired speech by connecting recorded speech data and text speech synthesis that can create any speech from text In the synthesizer,
A recorded voice database in which a plurality of voices having the same utterance content and different speech styles are stored as a selectable unit of voice fragments including at least one accent phrase;
A text analysis unit for identifying a recorded voice-compatible text in which the recorded voice stored in the recorded voice database can be used for the input text;
In the identified recorded speech-corresponding text, the speech component is extracted from the recorded speech database as a condition for selecting from a plurality of candidates the recorded speech that can be used for the recorded speech-corresponding text. The speech adjustment is determined by a statistical method, and the recorded speech is selected based on the speech tone component and the coincidence of the front and back phoneme environment, and the recorded speech is selected with priority as much as possible. And
In a text other than the identified recorded voice-corresponding text, a prosody generating unit that generates a prosody from the consistency with the speech tone component, fundamental frequency, speech speed, and power of the selected recorded voice;
A speech segment search unit that generates synthesized speech by text speech synthesis from the generated prosody;
A synthesis processing unit that performs synthesis processing as necessary to generate a smooth synthesized speech,
A voice synthesizer combined with recording / editing and text-to-speech synthesis, characterized by comprising:   A speech editing program using both recording and editing and text-to-speech synthesis, which is written in a computer-readable program language and causes the computer to function as a speech-synthesizing and text-to-speech synthesis combined speech synthesizer according to claim 1.   A recording medium comprising a computer-readable recording medium, wherein the recording / speech / synthesizing voice synthesis program according to claim 2 is recorded on the recording medium.

Images