JPH01284898A - Voice synthesizing device - Google Patents

Abstract

PURPOSE:To output a composite voice whose naturalness and articulation are both high by accumulating in advance a large quantity of waveforms as a dictionary and selecting and connecting the most suitable waveform to an input text. CONSTITUTION:An input text from a terminal 1 is analyzed by a text analyzing part 2, and based on an output of the analyzing part 2, meter information for synthesizing a voice is generated by a meter information generating part 3. On the other hand, a waveform dictionary 9 for storing a large quantity of waveform information at every appropriate unit for constituting an output voice such as a phoneme, etc., is provided, and by information from the analyzing part 2 and the generating part 3, an appropriate waveform is selected from the dictionary 9 by a waveform selecting part 8. In case when a desired waveform does not exist, a deformation is performed by a waveform deformation processing part 10 so as to conform to a use purpose with respect to the nearest waveform to the selecting condition, and in case when the desired waveform does not exist at all, a new waveform is generated by a waveform generating part 11. These waveforms from the selecting part 8, the processing part 10 and the generating part 11 are connected by a waveform connecting part 12.

Description

[Detailed Description of the Invention] "Industrial Application Field" This invention relates to a speech synthesis device that inputs text and outputs arbitrary speech according to the text. A rule-based synthesizer (-) that performs synthesis.

“Prior Art” Conventionally, in rule synthesis devices that output arbitrary speech, LPC (Linear Predictive) is used as a speech synthesis method.
onCorder) method, units such as CV, vcv, and cvc are set in consideration of the correspondence with phonemes and articulatory connections, and prosodic information such as fundamental frequency patterns is set using the accent form and mora in exhalation paragraphs. Many of them maintain a method of generating C2 independently from phonological information based on the number and other factors. However, these methods naturally use a sound source with a different fundamental frequency from that used during analysis during speech synthesis, which causes abnormal amplitude and spectrum Q due to the mismatch between the vocal tract spectrum and the sound source spectrum, which causes LPC parameters. This results in a reduction in the value, causing a deterioration in the quality of the synthesized sound. This is despite the fact that the LPC method is an analysis and synthesis method that assumes the independence of vocal tract parameters and sound source parameters. This degradation is considered to be a fundamental problem in using the LPC analysis synthesis method for ordered synthesis.

Another method is to describe the characteristics of speech in formants and specify the movement of formants (there is a method to obtain a regular synthesized sound from two methods, but automatic extraction of formants is difficult and the description of formant transitions is not sufficient, so LPC is used). Currently, the quality is not as good as the method used.

On the other hand, some methods have been proposed that avoid this problem and focus on original waveforms with high clarity.

However, in both cases, at most several types of waveforms are prepared for each phoneme or syllable, and the fundamental frequency and duration length are adjusted by truncating, repeating, or thinning out the waveforms. Therefore, synthetic speech cannot be controlled in detail, and has the disadvantage that it sounds like a combination of short sounds or a mechanical sound like a buzzer.

An object of the present invention is to provide a speech synthesis device that is capable of outputting synthesized speech with high naturalness and clarity in rule synthesis required for text synthesis.

"Means for Solving the Problem" According to the present invention, an input text is analyzed by a text analysis section, and based on the output of the text analysis section, prosody information for speech synthesis is generated by a prosody information generation section. On the other hand, the original waveform,
A waveform dictionary is provided that stores a large amount of waveform information including the phonological environment of utterance, fundamental frequency pattern shape, duration information, amplitude information, etc., and is more suitable than the waveform dictionary based on information from the text analysis section and the prosodic information generation section. If the desired waveform is selected in the waveform selection section and the desired waveform is not available, the waveform closest to the selection condition (on the other hand, the waveform modification processing section deforms it to match the purpose of use and the desired waveform is not found at all) If there is no waveform, a new waveform is generated by the waveform generation section.The waveforms from the waveform selection section, waveform modification processing section, and waveform generation section are connected at the waveform connection section.

In this way, according to the present invention, a large number of waveforms are stored as a dictionary, and output speech is synthesized by selecting and connecting the most suitable waveforms for the input text.
High clarity and natural sound can be obtained.

"Example" FIG. 1 is a block diagram showing one embodiment of the present invention.

That is, when text to be converted into speech is input from terminal l, the text analysis unit 2 performs morphological analysis such as dependency and part-of-speech analysis, kanji-kana conversion, and accent processing, and the phonological sequence buffer 7 and prosody information generation. Part 3 (2) Necessary information is sent.The information includes a symbol string indicating phoneme distinctions to the phoneme sequence buffer 7, the number of moras in an exhalation paragraph, the accent shape, and the prosodic information generator 3. The prosodic information generation unit 3 generates pitch patterns, duration patterns for each phoneme, etc. based on this information.
and amplitude patterns are generated according to the rules and written into the respective buffers 4, 5.6.

The waveform selection section 8 refers to the phoneme sequence buffer 7, pitch pattern buffer 41, time length buffer 5, and amplitude buffer 6, and selects the optimal waveform from the waveform dictionary M9. The waveform dictionary 9 has, for example, a configuration as shown in FIG. 2, and stores waveforms together with various information at the time of utterance. The various types of information include the phoneme type, the phoneme environment of about 7 degrees before and after, the average pitch within a phoneme, the slope of approximation by a linear straight line to indicate the shape of the pitch, the duration of the phoneme, and the center of the waveform. These are time length adjustment information indicating the start and end points of the number pitch, the RMS value (amplitude) of the normalized phoneme waveform, and the actual waveform data. The waveform dictionary 9 is created in advance by off-line processing based on a large amount of voice data. For example, approximately several hours of audio data, such as words and sentences uttered by a male announcer's name, are converted into AD at 12 KHz, and phonological labeling is performed by observing a digital tunagram. This audio data can be created by displaying a waveform of 20 to 30 m5 before and after the phoneme boundary of the label on a display and cutting it out with a cursor. As a general rule, the cutout position should be the O-intercept from the negative to the positive waveform, and rules should also be established (for example, cut out at the 0-intercept before the positive peak for each diphoneme).By doing this, it is possible to avoid errors at connection points. Continuity can be avoided and a smooth continuous waveform can be obtained.In addition, if the audio data is analyzed using LPC, etc. to extract the pitch, and phonemes that are similar in terms of pitch shape and duration are integrated, the extracted waveform can be obtained. The number of waveform dictionaries can be reduced and waveform dictionaries can be created efficiently.

The operation of the waveform selection section 8 will be described in more detail as shown in FIG. 3 as an example. First, a search phoneme sequence is set. Is the search phoneme series the corresponding phoneme? Set the center (= placement) by cutting out the input phoneme sequence from the input phoneme sequence by performing windowing using the number of environmental phonemes in the dictionary.If a waveform candidate is not found by searching the waveform dictionary 9, sequentially search the phoneme sequence from both sides. Search while deleting.Even if the search phoneme sequence contains only the corresponding phoneme, if no waveform candidates are found,
The waveform generator 11 generates a desired pitch waveform. Next, pitch conditions for the phonemes to be selected are set, taking into account the pitch pattern that is thought to have the greatest effect on the naturalness of the synthesized speech. This is determined based on the average pitch and pitch shape with reference to the pitch pattern buffer 4. Although the permissible range should be determined based on experimental values, it is considered that naturalness can be created if it is within approximately 5 inches of the desired value. Waveform 1
+ If ura are found, select them based on time length conditions. The time length condition is set from the time length of the time length buffer 5 and an allowable range determined from experimental values, similar to the pitch condition. If there is no waveform candidate that matches the time length condition, the waveform candidate closest to the condition is selected, and time length adjustment processing is performed at the waveform transformation processing unit 10 (=.If a waveform candidate is found, then the amplitude condition Make a selection.

In this case, as well as the time length condition, the amplitude of the amplitude buffer 6,
Conditions are set from the and permissible ranges. If a waveform candidate is not found, the waveform candidate closest to the condition (2) is selected, and amplitude adjustment processing is performed at the waveform modification processing section 10 (2).In this way, the waveform selection section 8 The waveform selected in , the waveform created in the waveform generation section 11 , and the waveform adjusted in the waveform modification processing section 10 are sent out to the waveform connection section 12 (''-), and are sequentially combined as an audio waveform to the output terminal 13 (= Output.

The waveform generation unit 11 generates a waveform with an arbitrary pitch using, for example, LPG technology. That is, LPC parameters indicating spectra corresponding to phonemes are stored, and a waveform is generated by driving pulses or residuals with a specified pitch. Although the use of LPG technology here is different from the purpose of the invention, this waveform generation unit 11 is considered to be a so-called rescue measure when there is no waveform at all, and is not used frequently.

The waveform modification processing section 10 also performs time length adjustment processing and amplitude adjustment processing. These processes will be explained below.

The time length adjustment process differs depending on whether the phoneme is a voiceless sound or a voiced sound. In the case of unvoiced sounds, if it is a plosive sound, it is dealt with by expanding or contracting the silent section, and if it is a fricative sound, the waveform is cut or used repeatedly so that the desired time length is reached from the center to the front and back. In the case of voiced sounds, the pitch position at the center of the waveform is stored in a dictionary of about 3 pitches, and if the waveform data is longer, they are thinned out, and if the waveform data is shorter, they are used repeatedly. .

The amplitude adjustment process refers to the amplitude value determined for each phoneme from the amplitude buffer, and calculates the RMS of the selected or generated waveform.
linearly adjust the amplitude value by the ratio of the two values.

"Effects of the Invention" As stated above, according to this invention (2), a large number of waveforms are stored as a dictionary, the most suitable waveforms are selected for the input text, and the output speech is synthesized by connecting them. As a result, it is possible to provide audio with high clarity, clarity, and naturalness.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a speech synthesis device according to the present invention, FIG. 2 is a diagram showing an example of the configuration of a waveform dictionary 9, and FIG. 3 is a diagram showing a method for selecting the most appropriate waveform from the waveform dictionary. It is a flow diagram. Patent applicant: Nippon Telegraph and Telephone Corporation Representative: Takashi Kusano Figure 3

Claims (1)

[Claims] (1) A speech synthesis device that outputs speech according to input text, including a text analysis section that analyzes the input text, and a prosody that generates prosody information for speech synthesis based on the output of the text analysis section. It stores a large amount of waveform information that describes the original waveform, phonological environment of utterance, fundamental frequency pattern shape, duration information, amplitude information, etc. for each appropriate unit for assembling output speech such as an information generation unit and phoneme. a waveform dictionary, a waveform selection unit that selects an appropriate waveform from the waveform dictionary based on information from the text analysis unit and the prosodic information generation unit, and a waveform selection unit that selects an appropriate waveform from the waveform dictionary; A waveform transformation processing unit that performs transformation to match the purpose of use, a waveform generation unit that generates a new waveform if the desired waveform is not found at all, and a waveform selection unit, waveform transformation processing unit, and waveform generation unit. A speech synthesis device comprising: a waveform connecting section that connects waveforms of.