JP2010008922A - Speech processing device, speech processing method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech processing device for generating high quality synthetic speech which is not artificial with less booming and buzzing noise. <P>SOLUTION: A phoneme selecting section 43 selects multiple phonemes for each segment of a phonological sequence of a target speech. A phoneme integrating section 44 generates an integrated phoneme by integrating multiple phonemes for each segment. A formant emphasis level estimating section 46 estimates a formant emphasis level by using at least either of a feature amount regarding the multiple phonemes and a feature amount regarding the integrated phonemes, for each segment. A formant emphasis filter section 45 emphasizes formant based on the emphasis level estimated for the integrated phoneme for each segment. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a voice processing device, a voice processing method, and a program.

  Artificially creating speech signals from arbitrary sentences is called text-to-speech synthesis. Text-to-speech synthesis is generally performed in three stages: a language processing unit, a prosody processing unit, and a speech synthesis unit.

  The input text is first subjected to morphological analysis and syntactic analysis in the language processing unit, and then subjected to accent and intonation processing in the prosody processing unit to obtain phoneme sequence / prosodic information (basic frequency, phoneme duration). Output). Finally, the speech synthesis unit synthesizes a speech signal from the phoneme sequence / prosodic information. Therefore, the speech synthesis method used for the speech synthesizer must be a method that can synthesize an arbitrary phoneme sequence generated by the prosody processing unit with an arbitrary prosody.

  Conventionally, as such a speech synthesis method, the input phoneme sequence / prosodic information is stored in advance for each of a plurality of synthesis units (synthesis unit sequences) obtained by dividing the input phoneme sequence. A speech synthesis method (unit selection type speech synthesis method) that synthesizes speech by selecting speech units from a large number of speech units and connecting the selected speech units between synthesis units. It has been known. For example, in the segment selection type speech synthesis method disclosed in Patent Document 1, the degree of speech synthesis degradation caused by speech synthesis is expressed by cost, and calculated using a predefined cost function. The speech segment is selected so that the cost to be played is reduced. For example, the distortion and connection distortion caused by editing and connecting speech units are quantified using cost, and based on this cost, a speech unit sequence used for speech synthesis is selected, and the selected speech unit is selected. Based on the single sequence, synthesized speech is generated.

  As in the speech synthesis method disclosed in Patent Document 1, in consideration of the degree of speech synthesis degradation caused by speech synthesis, an appropriate speech unit sequence is selected from a large number of speech units. Thus, it is possible to generate synthesized speech in which deterioration of sound quality due to editing and connection of speech segments is suppressed.

  However, the segment selection type speech synthesis method disclosed in Patent Document 1 has a problem in that the quality of synthesized speech is partially degraded. The reason for this is as follows.

  The first reason is that even if there are a large number of speech segments stored in advance, speech segments suitable for various phoneme / prosodic environments are not always present.

  The second reason is that an optimal unit sequence may not always be selected because the cost function cannot completely represent the degree of deterioration of the synthesized speech that a person actually feels.

  The third reason is that because there are so many speech segments, it is difficult to eliminate bad speech segments in advance, and it is also difficult to design a cost function for removing bad speech segments. This is because a defective speech unit may be suddenly mixed in the selected speech unit sequence.

  Therefore, instead of selecting one speech unit per synthesis unit, a plurality of speech units are selected per synthesis unit, and a new speech unit is generated by fusing them, thus generated. A method of synthesizing speech using speech segments has been disclosed (see Patent Document 2). Hereinafter, this method is referred to as a “multiple unit selective fusion type speech synthesis method”.

  In the multi-unit selective fusion type speech synthesis method disclosed in Patent Document 2, there is an appropriate speech unit suitable for the target phoneme / prosodic environment by fusing a plurality of speech units for each synthesis unit. If no optimal speech unit is selected, or if a defective unit is selected, a new high-quality speech unit can be generated, and this newly generated speech unit can be generated. By performing speech synthesis using fragments, it is possible to improve the above-described problems of the unit selection type speech synthesis method, and it is possible to realize speech synthesis with higher sound quality and higher stability.

  In this multi-unit selective fusion type speech synthesis method, the spectral envelope tends to be slightly dull compared to the original sound due to the side effect of averaging due to the fusion of speech units. is there. In order to subjectively improve such a feeling of being bulky or a buzzer, it is effective to apply a formant emphasis filter, which is often used in speech coding and speech synthesis, to the fused segments.

  The formant emphasis filter is a filter that outputs a speech waveform that emphasizes peaks and valleys due to the formant of the spectral envelope of the input speech waveform.If the formant can be emphasized to a reasonable degree, Buzzer feeling can be improved. In general, formant emphasis filters are adaptive in that they change the filter characteristics according to the spectral characteristics of the input waveform, but there is no objective measure for determining the appropriate degree of emphasis on how much formants are emphasized. However, it is often determined experimentally by subjective evaluation or the like, and is often controlled by specifying values such as hyperparameters from the outside.

For this reason, when used in the multi-unit selection type speech synthesis method, the degree of formant enhancement is determined experimentally by subjective evaluation or the like so that the subjective sound quality of the synthesized speech is improved overall. In other words, the same degree of formant emphasis is applied to all fused pieces.
JP 2001-282278 A JP 2005-164749

  However, the degree of blunting of the spectral envelope due to the fusion of speech elements usually varies depending on the synthesis unit and is not uniform. For example, if multiple segments selected for the synthesis unit have similar spectral envelopes, it is considered that the spectral envelope will not be dulled even if they are merged, but the position of the formant varies greatly between the segments, etc. If the spectral envelopes of the selected speech segments have different characteristics, there is a high possibility that the spectral envelope will become dull when they are merged.

  In such a situation, when formant emphasis filter with the same emphasis degree is uniformly applied to all speech segments, formant emphasis is insufficient at the place where the spectral envelope is greatly dull due to fusion. On the other hand, there is a problem that the formant is emphasized too much and the artificial sound is generated at the portion where the spectral envelope due to fusion is small.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a speech processing device, a speech processing method, and a program that can generate high-quality synthesized speech that is less artificial and has a low feeling of buzz and a buzzer. And

  The speech processing apparatus according to the present invention includes a first acquisition unit that acquires a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit, and prosodic information of each of the segments corresponding to the target speech A second acquisition unit that acquires a plurality of speech elements for each segment from a plurality of speech segments prepared in advance based on the prosodic information of the segment. A selection unit that selects a piece; a fusion unit that generates a fusion unit by fusing a plurality of the speech units selected for the segment for each segment; and each of the segments For each, at least one of a feature amount related to the plurality of speech units selected by the selection unit and a feature amount related to the fusion unit generated by the fusion unit is used. The estimation unit for estimating the degree of emphasis in formant enhancement to be performed on the fusion unit related to the segment, and the estimation unit for the fusion unit related to the segment for each segment. And a formant emphasis filter unit that performs formant emphasis based on the estimated degree of emphasis.

  According to the present invention, it is possible to generate a high-quality synthesized voice that is less artificial and has a buzzing feeling and is not artificial.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
A text-to-speech synthesizer (speech processor) according to a first embodiment of the present invention will be described.

  FIG. 1 shows an example of the overall configuration of a text-to-speech synthesizer (speech processor) that performs text-to-speech synthesis according to the present embodiment.

  As shown in FIG. 1, the text-to-speech synthesizer of this embodiment includes a text input unit 1, a language processing unit 2, a prosody processing unit 3, and a speech synthesis unit 4.

  The text input unit 1 inputs text.

  The language processing unit 2 performs morphological analysis and syntax analysis of the text input from the text input unit 1, and outputs the language analysis results obtained by the language analysis to the prosody processing unit 3.

  The prosodic control unit 3 inputs the language analysis result, performs accent and intonation processing from the language analysis result, generates a phoneme sequence and prosodic information, and outputs the generated phoneme sequence and prosodic information to the speech synthesis unit To do.

  The speech synthesizer 4 inputs the phoneme sequence and prosody information, generates a speech waveform from the phoneme sequence and prosody information, and outputs it.

  Hereinafter, the configuration and operation of the speech synthesizer 4 will be described in detail.

  FIG. 2 shows a configuration example of the speech synthesizer 4 of the present embodiment.

  As shown in FIG. 2, the speech synthesizer 4 includes a phoneme sequence / prosodic information input unit 41, a speech unit storage unit 42, a unit selection unit 43, a unit fusion unit 44, a formant enhancement filter unit 45, and formant enhancement. A degree estimation unit 46, a segment editing / connection unit 47, and a speech waveform output unit 48 are provided.

  A phoneme sequence / prosodic information input unit (hereinafter abbreviated as an information input unit) 41 receives phoneme sequence / prosodic information from the prosody control unit 3 as an input to the speech synthesis unit 4.

  A speech unit storage unit (hereinafter abbreviated as a unit storage unit) 42 stores a large amount of speech units. In addition, the unit storage unit 42 stores the phoneme / prosodic environment for each of the speech units together with the stored speech units.

  The unit selection unit 43 selects a plurality of speech units from the speech units stored in the unit storage unit 42.

  The unit fusion unit 44 unites a plurality of speech units selected by the unit selection unit 43 to generate a new speech unit (hereinafter also referred to as “fusion unit”).

  The formant emphasis filter unit 45 performs formant emphasis on the speech unit generated by the unit fusion unit 44 (that is, according to the degree of emphasis estimated by the next formant emphasis degree estimation unit 46) (that is, , Generate formant-enhanced fusion fragments).

  The formant emphasis degree estimation unit 46 estimates the degree of emphasizing formants in the formant emphasis filter unit 45.

  The segment editing / connecting unit 47 prosody modification and connection of speech units obtained from the formant emphasis filter unit 45 to generate a speech waveform of synthesized speech.

  The speech waveform output unit 48 outputs the speech waveform generated by the segment editing / connection unit 47.

  Note that the functions of the information input unit 41 to the voice waveform output unit 48 can be realized by a program stored in a computer.

  Next, each block of the speech synthesizer 4 in FIG. 2 will be described in detail.

<Information input part>
First, the information input unit 41 outputs the phoneme sequence / prosodic information input from the prosody control unit 3 to the segment selection unit 44. The phoneme sequence is a sequence of phoneme symbols, for example. The prosodic information includes, for example, a fundamental frequency, a phoneme duration, and power.

  Hereinafter, the phoneme sequence and prosody information input to the information input unit 41 are referred to as input phoneme sequence and input prosody information, respectively.

<Unit storage unit>
Next, a large amount of speech segments are stored in the segment storage unit 42 in units of speech used when generating synthesized speech (hereinafter referred to as “synthesis unit”).

  Here, the “synthesis unit” is a phoneme or a combination of phonemes (for example, semiphonemes), for example, semiphonemes, phonemes (C, V), diphones (CV, VC, VV), and triphones. (CVC, VCV), syllables (CV, V), and the like (where V represents a vowel and C represents a consonant), and may be variable length such as a mixture of these.

  Further, “speech segment” represents a waveform of a speech signal corresponding to a synthesis unit or a parameter series representing the characteristics thereof.

  FIG. 3 shows an example of speech segments accumulated in the segment storage unit 42. As shown in FIG. 3, in the segment storage unit 42, a speech unit that is a waveform of a speech signal of each phoneme is stored together with a unit number for identifying the speech unit. These speech segments are obtained by labeling a large number of separately recorded speech data for each phoneme and cutting out a speech waveform for each phoneme according to the label.

  The unit storage unit 42 stores a large amount of speech units and phoneme / prosodic environments corresponding to each speech unit.

  Here, the “phoneme / prosodic environment” is a combination of factors that become an environment for the corresponding speech segment. Factors include, for example, the phoneme name of the speech unit, the preceding phoneme, the subsequent phoneme, the subsequent phoneme, the fundamental frequency, the phoneme duration, power, the presence or absence of stress, the position from the accent core, the time from breathing, the utterance There are speed, feelings, etc.

  In addition to the above, the unit storage unit 42 also stores information used for selecting speech units among the acoustic features of the speech units, such as cepstrum coefficients at the start and end of the speech units.

  Hereinafter, the phoneme / prosodic environment and the acoustic feature amount of the speech unit stored in the unit storage unit 42 are collectively referred to as “unit environment”.

  FIG. 4 shows an example of the segment environment accumulated in the segment storage unit 42. In the environment storage unit 43 shown in FIG. 4, the unit environment is stored corresponding to the unit number of each speech unit stored in the unit storage unit 42. Here, as phoneme / prosodic environment, phonemes (phoneme names) corresponding to phonemes, adjacent phonemes (in this example, two phonemes before and after the phoneme), fundamental frequency, and phoneme duration are stored. A cepstrum coefficient at the beginning and end of a speech unit is stored as a feature amount.

  These segment environments are obtained by analyzing and extracting speech data from which speech segments are cut out. Further, FIG. 4 shows a case where the synthesis unit of a speech unit is a phoneme, but it may be a semiphoneme, a diphone, a triphone, a syllable, or a combination or variable length thereof.

<Unit selection unit>
Next, the operation of the speech synthesizer 4 in FIG. 2 will be described in detail.

  In FIG. 2, the phoneme sequence input to the segment selection unit 43 via the information input unit 41 is segmented by synthesis unit in the segment selection unit 47. Hereinafter, the divided synthesis unit is referred to as a “segment”.

  The segment selection unit 43 refers to the segment storage unit 42 based on the input input phoneme sequence and the input prosody information, and selects a combination of a plurality of speech units to be fused for each segment. To do.

  At this time, the unit selection unit 43 selects a combination of speech units to be merged so that distortion between the synthesized speech and the target speech when the speech is synthesized using each speech unit candidate is minimized. Here, similar to the general unit selection type speech synthesis method and the conventional multiple unit selection fusion type speech synthesis method, the unit selection unit 43 selects each speech unit candidate as a scale for selecting a speech unit. Using a cost that indirectly represents the magnitude of distortion between the synthesized speech and the target speech when the speech is synthesized using the speech, a combination of speech units to be merged is selected so that this cost is minimized.

  Here, the “target speech” refers to a (virtual) speech that is a target when synthesizing speech, that is, an ideal natural speech that realizes the input phoneme sequence and prosody. .

  First, the cost will be described.

  Costs representing the degree of distortion of the synthesized speech with respect to the target speech are roughly divided into two types of costs: target cost and connection cost.

  The target cost is a cost generated by using a speech segment (target segment) that is a cost calculation target in a target phoneme / prosodic environment.

  The connection cost is a cost that occurs when the target segment is connected to an adjacent speech segment.

  Specifically, it is as follows.

  The target cost includes distortion (basic frequency cost) caused by the difference (difference) between the fundamental frequency of the speech unit and the target fundamental frequency, and the difference (difference) between the phoneme duration of the speech unit and the target phoneme duration. ) Caused by the difference between the phoneme environment to which the speech segment belonged and the target phoneme environment (phoneme environment cost). Connection costs include distortion caused by spectral difference (difference) at the speech unit boundary (spectrum connection cost), distortion caused by fundamental frequency difference (difference) at the speech unit boundary (basic frequency connection cost), etc. There is.

  Any method may be used for selecting a plurality of speech segments per segment using cost.

  For example, the method disclosed in Patent Document 2 may be used. Here, the outline of this selection method will be described with reference to the case of selecting M speech units per segment with reference to the processing procedure example of FIG.

  First, in step S101, the segment selection unit 43 divides the input phoneme sequence into segments for each synthesis unit. Here, the number of divided segments is N.

  Next, in step S102, one speech segment sequence is selected for each segment from the speech segment group stored in the segment storage unit 42. In the selection at this time, based on the inputted target phoneme sequence / prosodic information and the information on the speech unit environment of the unit storage unit 42, the total cost (total cost) as a sequence is minimized. A speech unit sequence (optimal unit sequence) is obtained. The search for the optimum unit sequence can be efficiently performed by using dynamic programming (DP).

  Next, in step S103, an initial value “1” is set in the counter i representing the segment number.

  Next, in step S104, a cost is calculated for each of a plurality of speech element candidates for segment i. The cost used at this time includes the target cost of the speech unit candidate, the optimal speech unit of the segment before and after the speech unit candidate (the speech unit included in the optimal unit sequence), and the speech unit. The sum of the connection costs with the candidates is used.

  Next, in step S105, using the cost calculated in step S104, the top M speech elements with the lowest cost are selected for segment i.

  Next, in step S106, it is determined whether the counter i is N or less.

  If the counter i is less than or equal to N (YES in step S106), the process proceeds to step S107, the counter i is incremented by one, and then the process proceeds to step S104 to perform processing related to the next segment.

  If the counter i reaches N (NO in step S106), the segment selection process is terminated.

  As described above, the segment selection unit 44 selects M speech units for each segment and outputs the selected speech unit to the separation unit 45.

  The method for selecting a plurality of speech segments per segment in the segment selection unit 44 is not necessarily limited to the above-described method. Any method may be used as long as it can select an appropriate speech segment set.

<Fragment unit>
The unit fusion unit 44 fuses a plurality of speech units input from the unit selection unit 43 for each segment to generate a new speech unit.

  Any method may be used as a method for fusing speech segments.

  For example, the method disclosed in Patent Document 2 may be used. Here, this method will be described with reference to FIGS.

  FIG. 6 is a flowchart showing a procedure for generating a new speech waveform by fusing waveforms of a plurality of speech segments for one segment. FIG. 7 shows an example of generating a new speech unit (63 in the figure) by merging the unit combination candidates (60 in the figure) composed of three speech units selected for a certain segment. FIG.

  First, in step S201, a pitch waveform is cut out from each selected speech segment (for a certain segment).

  Here, the “pitch waveform” is a relatively short waveform having a length that is several times the basic period of speech, and has no fundamental period, and its spectrum represents the spectral envelope of the speech signal. Is.

  Any method may be used to extract such a pitch waveform, but one method is to use a fundamental period synchronization window. Here, the case where this method is used is used. Let's take an example.

  Specifically, a mark (pitch mark) is attached to the speech waveform of each speech unit at every basic period interval, and a Hanning window whose window length is twice the basic period centered on this pitch mark. A pitch waveform is cut out by windowing. A pitch waveform series 61 of FIG. 7 shows an example of a series of pitch waveforms obtained by cutting out from each speech unit of the unit combination candidate 60.

  Next, in step S202, the number of pitch waveforms is aligned so that the number of pitch waveforms for each speech unit is the same between speech units.

  At this time, the number of pitch waveforms to be aligned is the number of pitch waveforms necessary to generate a synthesized speech having a target phoneme duration length. good.

  A series with few pitch waveforms increases the number of pitch waveforms by duplicating several pitch waveforms included in the series, and a series with many pitch waveforms reduces the number of pitch waveforms by thinning out several pitch waveforms in the series. cut back. The pitch waveform series 62 in FIG. 7 shows an example in which the number of pitch waveforms is aligned to six.

  Next, in step S203, after arranging the number of pitch waveforms, a new pitch waveform sequence is generated by fusing pitch waveforms in the pitch waveform sequence corresponding to each speech unit for each position.

  For example, the pitch waveform 63 a included in the new pitch waveform 63 generated in FIG. 7 is obtained by fusing the sixth pitch waveforms 62 a, 62 b, and 62 c in the pitch waveform series 62. The new pitch waveform series 63 generated in this way is used as a fused speech unit.

  Here, as a method of merging pitch waveforms, for example, there are the following methods.

  The first method is simply a method of calculating the average of the pitch waveform.

  The second method is a method of averaging after correcting the position of each pitch waveform in the time direction so that the correlation between the pitch waveforms is maximized.

  The third method is a method in which the pitch waveform is divided into bands, the positions of the pitch waveforms are corrected and averaged so that the correlation between the pitch waveforms is maximized for each band, and the results are averaged between the bands.

  Any method may be used, but in this embodiment, the case of using the third method described last will be described as an example.

  Using the above-described method, the unit fusion unit 44 unites a plurality of speech units for each segment to generate a new speech unit, and outputs the new speech unit to the formant emphasis filter unit 45.

<Formant emphasis filter part>
Now, the speech waveform of the speech unit generated by the fusion as described above has a spectrum envelope that is less than the waveform of the speech unit of the fusion source due to the influence of the fusion, and some formants are weakened. As a result, the sense of clarity often decreases. Therefore, the formant emphasis filter unit 45 performs filtering for emphasizing the formant on the fusion unit input from the unit fusion unit 44 and outputs the result to the unit editing / connection unit 47.

  As the formant enhancement filter used here, for example, J. Chen et al. (J. Chen, etc., “Adaptive Postfiltering for Quality Enhancement of Coded Speech”, IEEE Trans. Speech and Audio Processing, vol. 3, Jan 1995 ) (Hereinafter referred to as Document 3) can be used.

  By applying such a formant emphasis filter to the speech waveform of the fusion unit, it is possible to emphasize the formant in the spectrum envelope and compensate for the decrease in clarity due to the fusion.

An outline of the formant emphasis filter will be described using the formant emphasis filter disclosed in Document 3 as an example. The formant emphasis filter disclosed in Document 3 is a filter having a transfer function as expressed by Equation (1).

However, P (z) is expressed by Equation (2). Here, a _i represents the i-th linear prediction coefficient (LPC) when the input waveform is subjected to linear prediction analysis, and M is the linear prediction order.

1 / [1-P (z / α)] in Equation (1) represents a linear prediction filter when α = 1, and has the same frequency response as the LPC spectrum of the input waveform. When α is made small, the frequency response becomes as if the LPC spectrum is dulled, and as it approaches 0, the frequency response becomes flat. Therefore, the frequency component with high power in the spectrum of the input waveform becomes larger, and the frequency component with low power becomes smaller, so that it has the effect of emphasizing peaks and valleys in the spectrum. In addition, since the negative envelope of a general speech spectrum envelope is seen from low to high, the frequency response of 1 / [1-P (z / α)] is generally the same. With a negative slope. That is, in addition to the effect of emphasizing the peaks and valleys of the spectrum, it has a low-pass characteristic as a side effect. Therefore, this low-pass characteristic is corrected by the terms [1-P (z / β)] and [1-μz ⁻¹ ]. [1-P (z / β)] is a filter having a zero point at the same frequency as the pole of the LPC spectrum, and has an effect of compensating for the inclination of the spectrum at 1 / [1-P (z / α)]. . On the other hand, [1-μz ⁻¹ ] is a term for adjusting so as to eliminate the inclination of the remaining spectrum with a simple high-pass filter. Note that G is a power adjustment gain for preventing the power from changing before and after filtering, and can be automatically determined according to the input waveform by the method disclosed in Document 3.

  In this formant emphasis filter, the degree of formant emphasis can be changed by changing the parameter α (however, it is necessary to determine appropriate β and μ that compensate for the low-pass characteristics according to the value of α). . As α is closer to 1, the degree of emphasis is stronger, and as α becomes smaller, the degree of emphasis becomes weaker, and when α is 0.5 or less, it is hardly emphasized. How much formant should be emphasized depends on the characteristics of the speech waveform. However, there is no objective measure to determine this, so the degree of formant enhancement is usually used when using a formant enhancement filter in speech coding or speech synthesis. Is obtained experimentally by subjective evaluation.

  However, in the multi-element selective fusion type synthesis method, the degree of blunting of the spectral envelope due to fusion can vary greatly from segment to segment. Therefore, if the same parameter is applied to the entire sentence or the like, the spectral envelope is greatly blunted by fusion. Whereas the formant emphasis is insufficient at the spot, the problem is that the formant is too weak and the artificial sound is generated at the spot where the spectral envelope due to fusion is small.

  Therefore, in the present embodiment, an appropriate formant enhancement degree is estimated by the formant enhancement degree estimation unit 46 for each speech unit formed by the fusion (or for each pitch waveform of each fusion unit). The formant emphasis filter unit 45 changes the coefficient of the formant emphasis filter according to the estimated formant emphasis degree. That is, the formant enhancement degree is adaptively controlled for each fusion unit (or for each pitch waveform of the fusion unit). Here, the formant emphasis degree given from the formant emphasis degree estimation unit 46 may change continuously from 0 (no emphasis) to 100 (strongest emphasis within the controllable range of the formant emphasis filter), for example. Alternatively, for example, it may be discrete such that it can be specified in five stages from 0 (no emphasis) to 4 (very strong emphasis). When the formant emphasis filter disclosed in the above-mentioned document 3 is used, when the formant emphasis degree estimated by the formant emphasis degree estimation unit 46 is large, the value of α is brought close to 1, and conversely, the formant degree is small. α is brought close to 0.5. The values of β and μ are also changed according to the value of α, but appropriate values of β and μ can be obtained experimentally for each α value.

  Also, mapping for specifically reflecting the formant enhancement degree estimated by the formant enhancement degree estimation unit 46 in the filter coefficient can be obtained experimentally by subjective evaluation or the like.

  In this embodiment, the case where the formant emphasis filter disclosed in Document 3 is used has been described. However, any formant emphasis filter can be used as long as the formant can be emphasized and the emphasis degree of the formant can be controlled by a parameter or the like. be able to.

<Formant emphasis degree estimation unit>
The formant emphasis degree estimation unit 46 generates an appropriate formant for the fusion unit based on the information of the fusion unit and the plurality of speech units of the fusion source given from the unit selection unit 43 and the unit fusion unit 44. The degree of emphasis is estimated, and the estimated formant emphasis degree is output to the formant emphasis filter unit 45.

  As described above, there is no objective scale that determines the appropriate formant enhancement level for a certain waveform, but the feature values related to the spectral envelope are compared between the fusion unit and the speech units of the fusion source. Thus, it is possible to estimate to some extent how much the spectral envelope has become dull due to the fusion of speech segments. Therefore, the formant emphasis degree estimation unit 46 estimates the dullness of the spectrum envelope due to the fusion by the following method, and determines the formant emphasis degree based on this.

  It is considered that the difference in the shape of the spectral envelope between each speech unit and the fusion unit becomes larger as the spectral envelope becomes dull due to the fusion. Therefore, if the difference in the shape of the spectrum envelope between each speech unit of the fusion source and the fusion unit can be estimated, it is considered that the dullness of the spectrum envelope due to the fusion of the speech units can be estimated.

  The parameters representing the characteristics of the spectrum envelope include cepstrum and LSP (Line Spectrum Pair). The following is an example of indirectly estimating the difference in the shape of the spectral envelope between each fusion unit and each fusion unit using the Mel frequency cepstrum coefficient (MFCC), which is one of the cepstrum. explain.

MFCC is an acoustic feature that is widely used in the field of speech recognition, and is often used as an evaluation measure of the above-mentioned “spectrum connection cost” in speech synthesis. The MFCC is a feature amount that takes into account human auditory characteristics, and has the advantage that the features of the spectral envelope can be expressed well even in a low dimension. The low order coefficients of the MFCC represent the shape of the spectral envelope, and the high order coefficients represent the details of the spectral envelope. Assuming that the i-th order MFCCs of the segment 1 and the segment 2 are c _1i and c _2i , respectively, the MFCC distance between the segment 1 and the segment 2 can be calculated by Equation (3).

  However, p represents the dimension of MFCC.

  In this example, the MFCC dimension is about 20th order.

  Next, a method of estimating the dullness of the spectrum envelope due to the fusion of speech units using the MFCC distance will be described.

  FIG. 8 shows an example of the processing procedure in this case.

  Here, it is assumed that the number of fusion elements that are the basis of the fusion element is N.

  First, the MFCC (c0) of the fusion unit is calculated (step S301).

Next, the counter i is initialized to 1 and D _sum is initialized to 0 (steps S302 and S303), and the process proceeds to step S304.

  In step S304, the MFCC (ci) of the i-th speech element among the N speech elements of the fusion source is calculated.

Next, the MFCC distance (D _i ) between c ₀ and c _i is calculated using Equation (3) (step S304).

In the next step S305, the calculated _{D i} by adding the _{D sum,} the process proceeds to step S307.

  In step S307, it is determined whether the counter i is N or less.

  If the counter i is equal to or smaller than N (YES in step S307), the process proceeds to step S308, the counter i is incremented by 1, and then the process proceeds to step S304 to perform processing related to the next speech unit. .

  If the counter i reaches N (NO in step S307), the loop process is terminated and the process proceeds to step S309.

In step S309, the average MFCC distance (D _mean ) is obtained by dividing D _sum by N, and all the processes are terminated.

  In the present embodiment, the average MFCC distance obtained in this way is used as an evaluation scale that reflects the dullness of the spectral envelope due to fusion. That is, the smaller the average MFCC distance, the smaller the spectral envelope dullness, and the larger the average MFCC distance, the larger the spectral envelope dullness. The average MFCC distance is used as it is or the distribution of the average MFCC distance. A value obtained by performing some conversion based on the above is defined as a dullness of the spectrum envelope.

  Next, it is necessary to obtain the formant emphasis degree based on the spectral envelope dullness obtained in this way, but it is thought that stronger formant emphasis should be applied as the spectral envelope dullness is greater, so here, The function is converted into a formant enhancement degree using a function that increases monotonously as the spectral envelope becomes dull (however, if the formant enhancement degree is a discrete value, it changes stepwise). As for the shape of the function, for example, it may increase linearly halfway with respect to the degree of bluntness of the spectrum envelope, and when it exceeds a certain threshold, it may take the upper limit of the formant enhancement degree, or it may be a sigmoid function. In addition, it may be of a shape such that the increase rate changes according to the degree of dullness of the spectrum envelope, and the parameters (slope, etc.) of those functions may be obtained experimentally.

  In this embodiment, as an example of a method for estimating the dullness of the spectrum envelope due to fusion, the method using the above MFCC has been described as an example. However, the acoustic parameters that can appropriately estimate the difference in the shape of the spectrum envelope are used. Any one may be used as long as it is present. For example, the square error of the LSP coefficient may be used, or the KL distance often used to calculate the difference in probability distribution by regarding the FFT spectrum obtained by FFT (Fast Fourier Transform) as a probability distribution. (Kullback-Leibler distance) may be calculated and used.

  Further, as a method for estimating the dullness of the spectrum envelope due to fusion, a method using the target cost calculated by the segment selection unit 43 is also conceivable. When a plurality of speech units of the fusion source are all selected from an appropriate phoneme / prosodic environment, it is considered that the target cost is reduced and the degree of blunting of the spectral envelope due to the fusion is also reduced. Conversely, if only speech segments that are different from the target phoneme / prosodic environment are selected, the target cost will increase, and the degree of blunting of the spectral envelope due to fusion will also increase. Therefore, it is considered that the target cost when the fusion source speech unit is selected may be used as one index indicating the degree of dullness of the spectrum envelope due to fusion. This method is indirect but much simpler than the method using the acoustic parameters described above.

  The formant emphasis degree estimation unit 46 outputs to the formant emphasis filter unit 45 the spectral envelope dullness caused by the fusion estimated as described above.

<Element editing / connection part>
The segment editing / connecting unit 47 generates a speech waveform of synthesized speech by deforming and connecting the speech units for each segment passed from the formant emphasizing unit 45 according to the input prosodic information.

  FIG. 9 is a diagram for explaining processing in the segment editing / connecting unit 47. In FIG. 9, speech units for the synthesis units of phonemes “a”, “N”, “s”, “a”, and “a” input from the formant emphasizing unit 45 are transformed and connected, and are referred to as “aNsaa”. The case where a speech waveform is generated is shown.

  In this example, a voiced speech segment is represented by a series of pitch waveforms. On the other hand, an unvoiced speech segment is represented as a waveform for each frame.

  A dotted line in FIG. 9 represents a segment boundary for each phoneme divided according to the target phoneme duration, and a white triangle represents a position (pitch mark) at which each pitch waveform arranged according to the target fundamental frequency is superimposed. .

  As shown in FIG. 9, for voiced sound, each pitch waveform of the speech unit is superimposed on the corresponding pitch mark, and for unvoiced sound, the waveform of each frame is pasted on the part corresponding to each frame in the segment. Then, a speech waveform having a desired prosody (here, fundamental frequency, phoneme duration length) is generated.

  As described above, according to the present embodiment, formant emphasis with an appropriate strength is performed for each segment in accordance with the formant bluntness caused by unit fusion, so that there is little feeling of vomit and buzzer and is not artificial. Sound quality synthesized speech can be generated.

(Second Embodiment)
A text-to-speech synthesizer (speech processing device) that performs text-to-speech synthesis according to a second embodiment of the present invention will be described.

  In the first embodiment, a large amount of calculation is required for the speech unit fusion processing and the formant emphasis processing. Therefore, the first embodiment may not be applicable to middleware applications with relatively low CPU specifications.

  Therefore, in this embodiment, a speech unit in which speech unit fusion and formant emphasis processing has been performed in advance is created offline, and an appropriate speech unit is created from the created speech unit during actual operation. A composite waveform is generated simply by selecting and connecting.

  The entire configuration example of the text-to-speech synthesizer according to this embodiment is the same as that in FIG. 1, and includes a text input unit 1, a language processing unit 2, a prosody processing unit 3, and a speech synthesis unit 4.

  FIG. 10 shows a configuration example of the speech synthesis unit 4 of the present embodiment.

  Hereinafter, with reference to FIG. 10, the present embodiment will be described focusing on differences from the first embodiment.

  As shown in FIG. 10, the speech synthesizer 4 of the present embodiment includes an information input unit 41, a segment storage unit 42, a segment selection unit 43, a segment editing / connection unit 47, and a speech waveform output unit 48. ing.

  Compared to the first embodiment (FIG. 2), the speech synthesizer 4 of this embodiment does not include the unit fusion unit 44, the formant enhancement filter unit 45, and the formant enhancement degree estimation unit 46 of FIG.

  In the segment storage unit 42 of the present embodiment, fused speech segments generated by a method described later are stored.

  The segment selection unit 44 of the first embodiment selects a plurality of speech segments for each segment, whereas the segment selection unit 44 of the present embodiment has one segment for each segment. Select the optimal sequence of speech units that have been merged.

  As the operation of the segment selection unit 44, for example, in comparison with the case where the flowchart of FIG. 5 is used in the first embodiment, in the present embodiment, only steps S101 and S102 in the flowchart of FIG. Good. Of course, the method of selecting the optimum sequence of fused speech segments for each segment is not limited to this, and various methods are possible.

  The operations of the segment editing / connecting unit 47 and the speech waveform output unit 48 are the same as those in the first embodiment.

  Next, a method for learning the fused speech unit stored in the speech unit storage unit 42 will be described with reference to FIGS. 11 and 12.

  In the present embodiment, a fused speech element creation unit 5 that creates a fused speech element is used. The fused speech element creation unit 5 may be included in the text-to-speech synthesizer of FIG. In this case, when the “formant-emphasized fused segment” for use in text-to-speech synthesis is generated, the speech synthesizer 4 in FIG. 1 may be replaced with the fused speech segment creation unit 5.

  Further, the fused speech segment creation unit 5 may not be included in the text speech synthesizer. In this case, for example, the fused speech segment creation unit 5 may be configured as an independent speech processing device (speech processing device that generates a “formant-emphasized fused segment” for use in text speech synthesis). . In this case, the independent speech processing apparatus may have a configuration in which the speech synthesis unit 4 in FIG. 1 is replaced with the fused speech unit creation unit 5.

  FIG. 11 shows a configuration example of the fused speech element creation unit 5.

  Since the configuration of the fused speech unit creation unit 5 is almost the same as the configuration of the speech synthesis unit 4 of the first embodiment, only the differences will be described here.

  The fused speech unit creating unit 5 has a speech unit output unit 49 instead of the unit editing / connecting unit 47 and the speech waveform output unit 48 of the speech synthesis unit 4 of the first embodiment. The segment editing / connecting unit 47 and the speech waveform output unit 48 of the speech synthesizer 4 of the first embodiment connect speech segments for each segment input from the formant emphasizing unit 45 to generate a synthesized waveform for the input text. The speech unit output unit 49 outputs the speech unit input from the formant emphasizing unit 45 as it is.

  That is, the fused speech unit creating unit 5 outputs the speech unit (formant-enhanced fused unit) to the speech unit storage unit 42 of the text-to-speech synthesizer in FIG. The fused segment) is stored in the speech segment storage unit 42.

  Next, a procedure for learning the fused speech unit to be stored in the speech unit storage unit 42 will be described.

  FIG. 12 shows an example of the processing procedure in this case.

  First, in step S501, a large number of sentences are input to the text-to-speech synthesizer provided with the fused speech unit creation unit 5 or an independent speech processing device.

  Next, in step S <b> 502, the fused speech element generated for each segment of each input sentence is output from the fused speech element generation unit 5.

  Next, in step S503, it is determined how many of the total number of speech units stored in the speech unit storage unit 42 designated from the outside are allocated to each unit type.

  Here, the unit type refers to a type classified by the phoneme environment of the speech unit. For example, the segment type / a / is a speech segment corresponding to the phoneme / a /.

  The number of segments allocated to each segment type is determined according to the appearance frequency of the speech segment of each segment type. For example, when the segment type / a / has a higher appearance frequency than the segment type / u / segment, a larger number of segments is allocated to the segment type / a /.

The number of speech units to be allocated to the segment type i and N _i.

  Next, in step S504, an initial value 1 is set to the segment type number i.

Next, in step S505, the united i speech units of unit type _i are extracted from the speech unit of unit type i output in step S502, the ones with the highest appearance frequency, N _{i at a} time.

  Next, in step S506, i is compared with the number of unit types.

  If i is less than or equal to the number of element types (YES in step S506), the process proceeds to step S507, the value of i is incremented by 1, and steps S505 to S506 are repeated.

  When i exceeds the number of segment types (that is, when processing for all segment types is completed) (NO in step S506), all processing is terminated.

  The fused speech unit extracted as described above is stored in the speech unit storage unit 42.

  Here, the number of speech units selected for storage in the speech unit storage unit 42 can be arbitrarily determined by a trade-off between the total speech unit size and the quality of the synthesized speech. If you select and store more speech units, the size will increase, but the quality of the synthesized speech can be increased, and if you reduce the number of speech units, the quality of the synthesized speech will be sacrificed, The size can be reduced.

  In addition, although the method of extracting the segment with high appearance frequency has been described above, it may be extracted using the feature amount of the speech unit such as a mel cepstrum calculated at both ends of the speech unit.

  In this case, the fused speech units output for each unit type are clustered using the feature values of the speech units, and the unit closest to the center (centroid) of each divided cluster is selected. Extract. The number of clusters in clustering is determined according to the number of segments allocated to each segment type.

  When extracting a segment based on the appearance frequency, an appropriate segment may not be extracted for a context with a low appearance frequency, and depending on the input text, the sound quality may be greatly degraded. When a segment is extracted by this method, a set of speech segments that covers the feature amount space as much as possible can be extracted, so that a more stable synthesized sound can be generated than when extracted based on the appearance frequency.

  As described above, according to the present embodiment, the process of fusing a plurality of speech units and the formant emphasis process are performed offline in advance, so that the calculation can be realized with a smaller calculation amount than the first embodiment, and the CPU specifications are compared. It can also be applied to low middleware applications.

  In addition, the total size of the speech segments to be stored can be determined in a scalable manner by a trade-off with the sound quality of the synthesized speech.

(Third embodiment)
A text-to-speech synthesizer according to a third embodiment of the present invention will be described.

  In the present embodiment, the estimation method of the formant enhancement degree estimation unit 46 is different from the example described in the first embodiment, and this difference will be mainly described below.

  In the first embodiment, as a method of estimating the formant emphasis degree by the formant emphasis degree estimation unit 46, a method of estimating the formant by calculating the difference in spectrum envelope between each fusion element and the fusion element. However, although there seems to be a high correlation between the difference in spectral envelope between each fusion source speech unit and the fusion unit and the bluntness of the spectral envelope due to fusion, There is no such relationship. Therefore, if there is a method that can directly determine the degree of bluntness of the spectrum envelope, it is considered possible to perform estimation with higher accuracy.

One method is to use a linear prediction pole (LP pole). The LP pole is a solution (complex number) obtained when (1-P (z)) = 0 for P (z) in Equation (2), and the position and unit of this solution on the z plane. From the relationship with the circle, the frequency and bandwidth of each formant can be estimated. Each pole is considered to correspond to each formant. For the i-th pole, when the angle of the line connecting the pole and the origin is θ _i and the distance between the pole and the origin is r _i , the frequency F _{i of the} i-th formant. And the bandwidth BW _i can be estimated as in Equation (4).

  If the frequency and bandwidth of each formant estimated in this way are used, it is considered that the degree of bluntness related to the formant in the spectral envelope can be estimated more accurately.

  Hereinafter, an example of a method of estimating the dullness of the spectrum envelope using the bandwidth of each formant estimated from the LP pole will be described with reference to FIG.

  FIG. 13 shows an example of a procedure for estimating the dullness of the spectral envelope using the bandwidth of each formant estimated from the LP pole.

  First, in step S601, the LP pole of the fusion unit is calculated. Specifically, an LPC analysis is performed on the speech waveform of the fusion unit, and (1−P (z)) = 0 for P (z) in Equation (2) having the obtained linear prediction coefficient as a coefficient. Get the solution when

  In the next step S602, the LP pole for each of the fusion source speech units is calculated by the same method as in step S601.

Next, in step S603, the sum of formant bandwidth ratios R _{sum is set} to 0, in step S604, the number of LP poles N _{LP used is set} to 0, and in step S605, the counter i is initialized to 1, respectively. The process proceeds to step S606.

  In step S606, it is determined whether or not the i-th LP pole of the fusion unit is on the real axis (that is, the imaginary term is 0). If it is on the real axis (YES in step S506), the process proceeds to step S620. After incrementing the value of the counter i by one, the process proceeds to S606 again.

  This is because the LP pole on the real axis does not correspond to the formant (contributes to the shape of the entire spectrum envelope), and therefore, on the real axis, the processing after step S607 is skipped and the LP pole corresponding to the formant. Only for consideration.

  If the LP pole is not on the real axis (NO in step S606), the process proceeds to step S607.

In step S607, after increasing the value of N _LP by 1, the process proceeds to step S608.

  In step S608, the formant bandwidth BWi for the i-th LP pole of the fusion unit is calculated using Equation (4).

In the next step S609, the bandwidth sum BW _{i_org_sum} relating to the formant of the fusion source speech unit is initialized to 0, and the process proceeds to step S610.

  In step S610, the counter k is initialized to 1, and the process proceeds to step S611.

In step S611, the k-th speech unit (referred to as “speech unit k”) among the fusion source speech units (total N _fused ) is the fusion unit among the LP poles of this speech unit. The LP pole corresponding to the formant represented by the i-th LP pole of the piece is selected. Specifically, among the LP poles of the speech unit k, the one closest to the i-th LP pole of the fusion unit is selected. For the distance between the LP poles, for example, Formula (5) (reference “Goncharoff, etc.,“ Interplation of LPC spectra via pole shifting. ”, IEEE ICASSP, Detroit, MI, Vol.1, pp.780-783, 1995 ”), where p _i is the complex number representation of the LP pole, r _i is the distance between the LP pole and the origin, and D (p0, p1) is the distance between the LP poles p ₀ and p ₁ . To express.

  Using this equation (5), the distance from the i-th LP pole of the fusion unit is calculated for each of the LP poles of the voice unit of the fusion source, and the LP pole with the shortest distance is selected.

In the next step S612, the bandwidth BW _{i_org_k} for the LP pole selected in step S611 is calculated using Equation (4).

Next, in step S613, BW _{i_org_k} calculated in step S612 is added to BW _{i_org_sum} .

Subsequently, in step S613, it is determined whether or not the counter k is equal to or less than the fusion source speech unit number N _fused .

If the counter k is equal to or less than N _fused (YES in step S613), the process proceeds to step S619, and after incrementing the value of the counter k by 1, the steps from step S611 are repeated. On the other hand, if the counter k exceeds N _fused (NO in step S613), the process proceeds to step S615.

In step S615, by dividing BW _{i_org_sum} by N _fused , an average value BW of formant bandwidths for the LP poles of each speech unit corresponding to the i-th LP pole of the fusion unit. _{i_org_mean} is calculated.

In the next step S616, the ratio of the bandwidth BWi of the i-th LP pole of the fusion unit to the BW _{i_org_mean} calculated in step S615 is added to the _sum Rsum of formant bandwidth ratios.

Subsequently, in step _S617 , it is determined whether the counter i is equal to or _smaller than a set value N _maxLP .

Here, N _maxLP represents the maximum value of the number of LP poles used to estimate the formant bluntness.

  This value is set to, for example, 1/2 of the analysis order in the LPC analysis.

If the counter i is _{equal to} or less than N _maxLP (YES in step _S617 ), the process proceeds to step S620, and after incrementing the value of the counter i by 1, the process from step S606 is repeated. On the other hand, when the counter i exceeds N _maxLP (NO in step _S617 ), the process proceeds to step _S618 .

In step S618, the average value R _mean of the formant bandwidth ratio is calculated by dividing the sum R _sum of the formant bandwidth ratios by the number N _LP of NP poles used, and all the processes are completed.

In the present embodiment, the average value R _mean of the formant bandwidth ratio calculated by the method as described above is used as a measure representing the dullness of the spectrum envelope due to the fusion of speech segments. This value is close to 1 when the bandwidth of the formant is not substantially changed and the spectral envelope is hardly dull, and is 1 when the bandwidth of the formant is wider than the speech unit of the fusion source and the spectral envelope is dull. It is considered that the larger the value is, the larger the dullness of the spectrum envelope becomes. Therefore, in the present embodiment, the formant emphasis degree is calculated from this R _mean by using some function such that no emphasis is given when R _mean is 1 or less, and the emphasis degree monotonously increases when R _mean is greater than 1. I will do it.

  In this way, by using the formant bandwidth estimated for the fusion unit and each speech unit of the fusion source, the degree of bluntness of the spectrum envelope due to fusion is more than that when using the difference in spectrum envelope shape. Since it is obtained directly, it is possible to estimate the formant emphasis degree with higher accuracy.

(Fourth embodiment)
A text-to-speech synthesizer according to a fourth embodiment of the present invention will be described.

  In the present embodiment, the estimation method of the formant emphasis degree estimation unit 46 is different from the examples described in the first and third embodiments. Hereinafter, this difference will be mainly described.

In the third embodiment, the degree of dullness in the entire spectrum envelope is estimated by averaging the formant bandwidth ratio obtained for each LP pole of the fusion unit. It may be possible that the dullness varies depending on the formant. Therefore, by using the formant bandwidth ratio (hereinafter referred to as R _i ) obtained for each LP pole as it is, it is possible to perform formant enhancement in which the degree of enhancement differs for each formant.

Here, assuming that the i-th LP pole of the fusion unit is p _i , P (z) in Expression (2) can be expressed as Expression (6).

When the prediction residual when the LPC analysis is performed is input to a filter (linear prediction filter) having a transfer function of H (z) = 1 / (1-P (z)), the original waveform can be completely reproduced. When the p _i inputting the prediction residual in the filter has been changed so as to approach the unit circle on the z plane, narrowed formant bandwidth corresponding to the i-th LP pole, consequently, to emphasize this formant Can do. That is, if a filter in which p _i is appropriately changed according to R _i is used as a formant emphasis filter, it is possible to perform appropriate formant emphasis for each formant.

  FIG. 14 shows a configuration example of the formant emphasis filter unit 45 of this embodiment.

  The LPC analysis unit 451 performs LPC analysis on the input waveform, and outputs the calculated LPC to the LPC deformation unit 452 and the prediction residual to the linear prediction filter unit 453.

The LPC deforming unit 452 modifies the LPC coefficient according to the formant bandwidth ratio R _i for each LP pole input from the formant enhancement degree estimating unit 46, and gives the deformed LPC coefficient to the linear prediction filter unit 453.

  The linear prediction filter unit 453 outputs a waveform with formant emphasis by filtering the prediction residual input from the LPC analysis unit 451 using the LPC coefficient given from the LPC deformation unit 452 as a filter coefficient.

Note that the LPC deformation unit 452 first obtains the mathematical expression P (z) from the input LPC coefficient, and then factorizes (1-P (z)) as shown in the mathematical expression (6) to obtain the LP pole. Get p _i .

Next, the LP pole p _i is changed according to R _i .

For example, if changed to Equation (7), the bandwidth of the formant becomes 1 / R _i times, and the bandwidth can be narrowed so as to approach the average formant bandwidth of the speech unit of the fusion source. It is.

A modified LPC coefficient can be obtained by substituting the LP pole p _i changed according to R _i in this way into Equation (6) and expanding the equation.

  In the present embodiment, the method of changing the degree of emphasis for each formant using the LP pole obtained for the fusion unit and each speech unit of the fusion source has been described. However, in addition to this method, for each formant or frequency It is also possible to perform formant emphasis that the degree of emphasis varies depending on the band.

  For example, the formant emphasis degree estimation unit 46 divides the waveform of the fusion unit and each speech unit of the fusion source into a plurality of frequency bands, and estimates the dullness of the spectrum envelope in each band, thereby obtaining each band. Estimate the degree of formant emphasis at. Then, after the formant emphasis filter unit 45 performs formant emphasis according to the emphasis degree of each band input from the formant emphasis degree estimation unit 46 with respect to the waveform of each frequency band obtained by dividing the waveform of the fusion unit into bands, If the waveforms are added between the frequency bands, it is possible to perform spectrum enhancement in accordance with the degree of spectral envelope dullness in each frequency band.

Each of the above functions can be realized even if it is described as software and processed by a computer having an appropriate mechanism.
The present embodiment can also be implemented as a program for causing a computer to execute a predetermined procedure, causing a computer to function as a predetermined means, or causing a computer to realize a predetermined function. In addition, the present invention can be implemented as a computer-readable recording medium on which the program is recorded.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

The block diagram which shows the structural example of the text speech synthesizer which concerns on one Embodiment of this invention. The block diagram which shows the structural example of the speech synthesizer which concerns on the embodiment The figure which shows the example of the speech unit accumulate | stored in the speech unit memory | storage part which concerns on the embodiment The figure which shows the example of the segment attribute information accumulate | stored in the speech unit memory | storage part which concerns on the embodiment Flow chart showing an example of a procedure for selecting speech segments Flow chart showing an example of a procedure for generating a new speech waveform by fusing speech waveforms The figure for demonstrating the example which unites the unit combination candidate which consists of three selected speech units, and produces | generates a new speech unit Flow chart showing an example of a procedure for estimating the dullness of the spectral envelope due to the fusion of speech segments The figure for demonstrating the process in the segment edit and the connection part which concerns on the same embodiment The block diagram which shows the other structural example of the speech synthesizer which concerns on the same embodiment The block diagram which shows the structural example of the united speech unit preparation part which concerns on the same embodiment Flow chart showing an example of a procedure for learning a fused speech unit Flow chart showing an example of a procedure for estimating the formant emphasis degree The block diagram which shows the structural example of the formant emphasis filter part which concerns on the same embodiment

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Text input part, 2 ... Language processing part, 3 ... Prosody processing part, 4 ... Speech synthesis part, 41 ... Phoneme series / prosodic information input part, 42 ... Speech unit memory | storage part, 43 ... Unit selection part, 44 ... unit fusion unit, 45 ... formant emphasis filter unit, 46 ... formant emphasis degree estimation unit, 47 ... unit editing / connection unit, 48 ... speech waveform output unit, 49 ... speech unit output unit, 5 ... fused speech Segment creation unit, 451... LPC analysis unit, 452... LPC deformation unit, 453... Linear prediction filter unit

Claims (15)

A first acquisition unit that acquires a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit;
A second acquisition unit that acquires prosodic information of each of the segments corresponding to the target speech;
For each segment, for each segment, a selection unit that selects a plurality of speech units from a plurality of speech units prepared in advance based on the prosodic information of the segment;
For each of the segments, a fusion unit that generates a fusion unit by fusing a plurality of the speech units selected for the segment;
For each of the segments, using at least one of a feature amount related to the plurality of speech units selected by the selection unit and a feature amount related to the fusion unit generated by the fusion unit, the segment An estimation unit for estimating an enhancement degree in formant enhancement to be performed on the fusion unit according to
An audio processing apparatus comprising: a formant emphasis filter unit that performs formant emphasis based on the degree of emphasis estimated by the estimation unit for each fusion segment related to the segment for each segment.   The generator further includes a generating unit that generates a synthesized speech based on a speech waveform related to the form fragment emphasized by the formant emphasizing filter unit obtained for each segment by the formant emphasizing filter unit. Item 6. The speech processing apparatus according to Item 1.   The speech processing apparatus according to claim 1, further comprising an output unit that directly outputs the fusion element that has been subjected to formant emphasis obtained by the formant emphasis filter unit for each of the segments.   The speech processing apparatus according to claim 3, wherein the output unit outputs the fusion unit to a storage unit that stores a speech unit for use in text-to-speech synthesis.   5. The speech processing apparatus according to claim 1, further comprising a speech unit storage unit that stores the plurality of speech units prepared in advance.   The estimation unit estimates, for each of the segments, how blunt the spectrum envelope of the fusion unit generated by the fusion unit is from the spectrum envelope of the speech unit selected by the selection unit. 6. The speech processing apparatus according to claim 1, wherein a stronger formant emphasis degree is estimated for a segment having a greater degree of bluntness of the estimated spectral envelope.   The estimation unit estimates a difference between a spectrum envelope of the fusion unit generated by the fusion unit and a shape of a spectrum envelope of the speech unit selected by the selection unit for each segment. The speech processing apparatus according to claim 1, wherein a stronger formant enhancement degree is estimated for a segment having a larger difference in estimated spectrum envelope shape.   The estimation unit selects, for each segment, a difference between a target speech and a speech generated by the fusion unit generated by the fusion unit, by the prosody information corresponding to the target speech of the segment and the selection unit. 2. The formant emphasis degree is estimated for a segment that is estimated from the prosodic information of the speech segment and has a larger difference between the estimated target speech and the speech of the fusion segment. 6. The sound processing device according to any one of 5 above. The estimation unit estimates the formant emphasis degree for each of the plurality of segments for each formant or for each frequency band divided into a plurality of segments,
6. The formant emphasizing filter unit performs formant emphasis having different strengths between formants or frequency bands in accordance with the formant emphasis degree estimated for each formant or frequency band. The voice processing device according to claim 1. A speech processing method of a speech processing apparatus including a first acquisition unit, a second acquisition unit, a selection unit, a fusion unit, an estimation unit, and a formant emphasis filter unit,
The first obtaining unit obtaining a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit;
The second obtaining unit obtaining prosodic information of each of the segments corresponding to the target speech;
The selection unit selects a plurality of speech units from a plurality of speech units prepared in advance for each segment based on the prosodic information of the segment for each segment. When,
The fusion unit for each segment, generating a fusion unit by fusing the plurality of speech units selected for the segment;
For each of the segments, the estimation unit calculates at least one of a feature amount related to the plurality of speech units selected by the selection unit and a feature amount related to the fusion unit generated by the fusion unit. Using to estimate the degree of emphasis in formant emphasis to be performed on the fusion unit related to the segment;
The formant emphasizing filter unit includes, for each segment, performing formant emphasis based on the degree of emphasis estimated by the estimation unit for the fusion unit related to the segment. Processing method. The voice processing device further includes a generation unit,
In the speech processing method, the generation unit generates a synthesized speech based on a speech waveform relating to the fusion element obtained by the formant emphasis filter unit for each segment and obtained by the formant emphasis filter unit. The speech processing method according to claim 10, further comprising: The voice processing device further includes an output unit,
The speech processing method further includes a step in which the output unit directly outputs the form fragment emphasized by the formant emphasis filter unit obtained by the formant emphasis filter unit for each of the segments as it is. The voice processing method described in 1. A program for causing a computer to function as a speech processing apparatus including a first acquisition unit, a second acquisition unit, a selection unit, a fusion unit, an estimation unit, and a formant emphasis filter unit,
The first obtaining unit obtaining a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit;
The second obtaining unit obtaining prosodic information of each of the segments corresponding to the target speech;
The selection unit selects a plurality of speech units from a plurality of speech units prepared in advance for each segment based on the prosodic information of the segment for each segment. When,
The fusion unit for each segment, generating a fusion unit by fusing the plurality of speech units selected for the segment;
For each of the segments, the estimation unit calculates at least one of a feature amount related to the plurality of speech units selected by the selection unit and a feature amount related to the fusion unit generated by the fusion unit. Using to estimate the degree of emphasis in formant emphasis to be performed on the fusion unit related to the segment;
A program for causing the computer to execute, for each of the segments, the step of performing formant emphasis based on the degree of enhancement estimated by the estimation unit for the fusion unit associated with the segment for each of the segments. . The voice processing device further includes a generation unit,
The program further includes a step in which the generation unit generates a synthesized speech based on a speech waveform related to the fusion element that is formant-emphasized obtained by the formant emphasis filter unit for each of the segments. The voice processing apparatus according to claim 13, wherein the voice processing apparatus is executed by a computer. The voice processing device further includes an output unit,
The computer program further causes the computer to execute a step of outputting the fusion fragment, which is obtained by the formant emphasis filter unit for each of the segments, as it is, and outputs the fusion fragment as it is. 14. The voice processing device according to 13.

Images