JP3368948B2 - Voice rule synthesizer - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech rule synthesizer, and more particularly to a technique for creating a power envelope in a speech rule synthesizer. 2. Description of the Related Art Conventionally, in a power control method in speech rule synthesis, a sound source amplitude pattern generation unit only performs processing to such an extent that power patterns synchronized with a spectrum pattern are smoothly combined and deformed. That is, a method is generally used in which the power of one point at the center position of a vowel is derived using a rule, and the natural utterance power pattern synchronized with the spectral pattern is smoothly deformed by appropriately interpolating between the center positions of the vowels. It was a target. Of course, in speech rule synthesis, there are very few examples in which power method control has been studied in comparison to techniques such as spectrum control. However, as a result of recent research and investigation, it has become clear that in the power control method, the shape of the power envelope is not irrelevant to other perceptual factors such as the duration (Hamagami et al .: " Effect of Power Envelope on Hearing ", Proc. Of the Acoustical Society of Japan, November 1992
Month). On the other hand, a rule-based speech synthesis method that performs duration control based on the vowel part energy centroid position (Kato: "Japanese rhythm rule focusing on vowel part energy centroid position" IEICE Technical Report (1993) May 22
When using day), the position of the energy center of gravity and the shape of the power envelope are directly correlated, so that it becomes necessary to control the power envelope according to the duration rule. [0004] From these viewpoints,
A new model for the shape of the power envelope is needed.
Further, what is desired in considering a power control model used for rule synthesis is to obtain a model capable of directly controlling a target physical characteristic by a rule. SUMMARY OF THE INVENTION It is an object of the present invention to create a natural power envelope and improve the quality of synthesized speech in a speech rule synthesis device. [0005] In order to realize high-quality regular speech synthesis, the present inventor has proposed a power pattern generation process using a high-accuracy model taking into account the effect on hearing. And applied it to the power pattern. Because the improvement of the conventional speech synthesis method was applied mainly to important parameters such as spectrum and pitch, the shape of the power envelope, which had not been regarded as important, can be ignored as a cause of the quality degradation of the synthesized sound. Because it is gone. The phoneme duration generation unit in the general speech rule synthesis method determines the duration of each phoneme based on the position of the center of gravity of the energy of the vowel, and the outline of the power envelope itself is the phoneme. Since it has been confirmed that it has an effect on the duration sensation, it is necessary to develop a power envelope model that can accurately reflect the excellent duration rule in synthesized speech. First, the formation of the power envelope model of the present invention will be described based on a naturally uttered power envelope.
FIG. 4 shows an audio power envelope when “explosive sound is made”. The horizontal axis represents each voice, and the vertical axis represents the peak power of the voiced sound block in the text normalized by the peak power in the text. The envelope connecting them is the outline of the macro envelope. When the same utterance is made several times, the outline of the envelope normalized by the average value of the power falls within a certain range of variation (± 4 db), and the power for naturalness during utterance is reduced. It was found that it coincided with the fluctuation detection limit. That is, during spontaneous utterance, the existence of a fixed rough power profile, which originally has some variation within the detection limit, is recognized. This is the macro envelope. The macro-envelope is a power change of a voiced sound using a vocal cord vibration as a sound source that generates a large power stably like a vowel during utterance. On the other hand, there are places where sharp power changes characteristic of consonants occur. This short-time power change is generated in synchronization with the consonant spectrum change due to rupture or friction of the lips or the like. This is called a microenvelope. Microenvelopes have a crucial role in the phonological properties of consonants. According to the present invention, based on the above principle, in order to achieve the above object, in a speech rule synthesizing apparatus, a power control unit controls a gradual change in speech power characterized by voiced sounds including vowels. A database is created in which a model and a micro-envelope model that controls fine power changes characterized by articulation of consonants are stored for each vowel-consonant-vowel combination, and a speech power envelope is formed using this database. According to the power control method of the present invention, a representative pattern having a rough outline formed by connecting the central powers of vowels of a natural power envelope with straight lines is used. A change of about ± 4 db from this pattern does not cause auditory discomfort. As a result, a rough outline (macro pattern) connecting the central powers of vowels can be represented by a small parameter called vowel central power. On the other hand, sharp power changes (micro patterns) characteristic of consonants can be handled separately from macro patterns in synchronization with the consonant spectral parameters. By expressing the power pattern of the voice to be synthesized as a macro change and a micro change using a center power of a vowel and a normalized power pattern stored in a database in synchronization with the spectrum, a natural change is obtained. Synthesized speech can be obtained. In addition, when performing speech rule synthesis, a necessary model and a control method thereof can be established in order to determine a power envelope by a rule. It works. From the power envelope of a naturally uttered sentence,
As a result of obtaining and re-synthesizing the macro-envelope and the micro-envelope using the above-mentioned separation method, it was confirmed that an envelope comparable to the power pattern of natural speech was obtained. It was also confirmed that the speech synthesized using the power envelope is indistinguishable from the synthesized speech synthesized using the natural power envelope. Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration of a Japanese speech synthesis system according to an embodiment of the present invention. The text input device 1 inputs a synthesized text to be synthesized. The text analysis unit 2 analyzes the synthesized sentence using the pronunciation information generation rule 3 and converts it into a phoneme symbol string to which pronunciation information such as accent information, pause, and vowel de-voice necessary for speech synthesis processing is added. I do. The phoneme duration generation unit 4 determines the phoneme duration of the phoneme symbol string generated by the text analysis unit 2 according to the rhythm rule 5. The sound amplitude pattern generation unit 6 determines the power envelope of the voice according to the power rule 7. The pitch pattern generation unit 8 determines a point pitch pattern for each accent phrase from the prosodic control rules 9 and generates a continuous point pitch pattern by interpolating them. The sound source generator 10 generates a sound source based on the power pattern and the pitch pattern. The spectrum pattern generation unit 11 uses the phonological improvement rule 12
The speech synthesis basic unit database 13 is searched from the types of phonemes such as vowels and consonants, and the spectra of the phonemes are combined according to the phoneme combination rules 14 to form a formant pattern. The speech synthesizer 15 creates a synthesized speech from the sound source information obtained from the sound source generation unit 10 and the formant pattern obtained from the spectrum pattern generation unit 11. The generated synthesized voice is uttered outside by the speaker 16. Next, a power control method of the sound source amplitude pattern generation unit 6, which is a feature of the present invention, will be described. First, the separation method of the macro envelope will be described with reference to FIG. FIG. 2 shows a part of the natural audio power envelope 21. In the figure, the horizontal axis represents time, V represents the time when a vowel is uttered, and C represents the time when a consonant is uttered. The vertical axis represents power in dB. First, for each vowel V, V, the vowel energy centroid (CEGV) time given by the rhythm rule 5 is determined as an anchor point. Next, a power change between CEGVs is synthesized according to a consonant between CEGVs. For that purpose, the concept of a consonant target (TC) is introduced.
The TC is composed of two parameters, “division coefficient α” and “dip coefficient β”, and is stored in a database in advance for each VCV (vowel-consonant-vowel). The division coefficient α represents the point at which the power becomes minimum in the consonant duration in the VCV by the internal division ratio of the consonant duration. When viewed from the line connecting the two CEGVs of the dip coefficient β and the VCV, the degree of drop at the point where the power becomes the smallest in the consonant section is expressed in dB. Macro envelope 22
Starts from the preceding CEGV, passes through TC in the form of a cosine function, and reaches the following CEGV. A power change accompanying a change in the articulation of the VCV is controlled by α and β.
As described above, the macro envelope 22 models a smooth change in voiced sound power in the process of passing from a vowel to the consonant to the next vowel. Next, the micro envelope will be described with reference to FIG. The micro-envelope deals with a relatively fast-changing power pattern such as a consonant sound source that cannot be expressed by the macro-envelope 22 alone. The microenvelope is obtained by time-stretching the normalized microenvelope in the database. The method of creating a normalized microenvelope that is made into a database in advance is as follows. The macro envelope 22 is obtained from the naturally uttered VCV power envelope 31 by the above-described method. The power 31 is re-analyzed in the consonant section of the VCV waveform with a window corresponding to the waveform characteristics of the consonant. From the re-analysis power envelope 32, the macro envelope 2
2 is subtracted to obtain a difference power 33. The consonant section of the differential power is compressed to a fixed length of about a fraction to obtain a normalized micro envelope 34. Here, it is compressed to 10 samples, and the normalized micro envelope 34 is obtained. The micro envelope 34 is parameterized for each VCV combination. Next, a method of generating an amplitude pattern by the sound source amplitude pattern generation unit 6 in FIG. 1 will be described. In the power rule 6, a database of VCV combination utterance patterns is stored as a basic unit of speech used for rule synthesis. A consonant target parameter and a normalized micro-envelope are obtained for each VCV using the above-described macro-envelope and micro-envelope separation methods. At the vowel part energy center of gravity given by the phoneme duration generator 4, the vowel is determined by the type of the vowel, the point pitch frequency given by the prosody control rule at that time, the in-sentence segment number, and the in-segment mora number. The center of gravity power (Pv) is determined. For the consonant C between the two vowels V1 and V2 having the powers of Pv1 and Pv2, a consonant target and a normalized microenvelope are searched and expanded and contracted to fit within a given duration, and the macroenvelope is
As a result of creating and adding the microenvelopes, a combined power method is obtained. As described above, according to the present invention,
By using a model that independently treats a macro envelope that strongly affects the naturalness of a synthesized sound and a micro envelope that affects phonology, a power envelope close to the shape of natural speech can be easily generated. Thereby, the quality of the synthesized speech can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram showing a configuration of a speech synthesis system according to an embodiment of the present invention. FIG. 2 is a view for explaining how to obtain a macro envelope used for the power rule in FIG. 1; FIG. 3 is a view for explaining a method of obtaining a micro envelope used for the power rule in FIG. 2; FIG. 4 is a graph for explaining a variation of a macro envelope in natural utterance. [Description of Signs] 1 ... Sentence input device 2 ... Text analysis unit 3 ... Pronunciation information generation rule 4 ... Phonological duration length generation unit 5 ... Rhythm rule 6 ... Sound amplitude pattern generation unit 7 ... Power rule 8 ... Pitch pattern generation unit 9 Prosody control rule 10 Sound source generation unit 11 Spectrum pattern generation unit 12 Phonological improvement rule 13 Speech synthesis basic unit database 14 Phoneme connection rule 15 Speech synthesizer 16 Speaker 21 Natural sound power envelope 22 ... macro envelope 31 ... VCV power envelope 32 ... reanalysis power envelope 33 ... difference power 34 ... normalized micro envelope

Claims (1)

(57) [Claims 1] In the speech rule synthesizing device, a power control unit controls a gradual change of speech power characterized by voiced sounds including vowels, and a macro envelope model. A microenvelope model for controlling a fine power change characterized by articulation of a consonant is provided for each vowel-consonant-vowel combination, and a speech power envelope is formed using this database. Voice rule synthesizer.