JP5075865B2 - Audio processing apparatus, method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech processing device, a method, and a program which integrate an elementary speech unit without breaking features of a sound source and a vocal tract filter inherent in a speech waveform. <P>SOLUTION: A phoneme-metric input reception section 41 receives inputs of a plurality of segments obtained by dividing phoneme series corresponding to target voice by a synthetic unit and metrical information corresponding to each of the plurality of segments. An acquisition section 43 acquires a plurality of elementary speech units related with the segment and the metrical information for each of the plurality of segments. A vocal tract filter component integration section 45 integrates a vocal tract filter component of the plurality of acquired elementary speech units for every segment. A sound source component integration section 46 expands and contracts a sound source component of a periodic component of the plurality of acquired elementary speech units based on a fundamental frequency or a shape of a waveform of the sound source component and integrates them for every segment. An elementary unit integration section 44 integrates the plurality of acquired elementary speech units for every segment by filtering the integrated sound source component using the vocal tract filter. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to an audio processing apparatus, method, and program.

  In recent years, in speech synthesizers that artificially generate speech signals from arbitrary sentences, improvement in sound quality is required.

  For example, in Patent Document 1, in order to improve the problem that the sound quality of the synthesized sound partially deteriorates when there is no appropriate speech unit, a plurality of speech units are selected per synthesis unit, A method (multiple segment selection and fusion method) for synthesizing speech by generating a new speech unit by fusing for each synthesis unit is disclosed.

  In Non-Patent Document 1, when a plurality of speech segments are fused, they are fused by dividing into a periodic component (periodic component) and an aperiodic component (non-periodic component). Further, it is disclosed that the feature amount relating to the sound source and the feature amount relating to the vocal tract filter are divided into the respective feature amounts.

Japanese Patent Laid-Open No. 2005-164749

Akihiro Morita and Takehiko Tsujishima, “Speech synthesis by multiple unit selection and fusion method considering non-periodic components in voiced sound”, Acoustical Society of Japan Spring Meeting, 295-296, 2008

  However, the speech unit fusion method disclosed in Patent Document 1 is basically a method of averaging a plurality of speech waveforms, and various features related to speech generation processes such as sound source and vocal tract filter characteristics. It fuses things that are mixed with various features. For this reason, it is not clear how the effect of the fusion appears in which feature, and as a result, the feature of each component in the speech waveform may be destroyed by the fusion, and the sound quality may be deteriorated.

  Note that Non-Patent Document 1 discloses a method in which non-periodic components are divided into feature quantities of a sound source and a vocal tract filter and fused together. However, in non-periodic components driven by a noisy sound source, It is not necessary to consider the shape of the sound source waveform itself, as long as the power changes with time and the frequency characteristics only need to be expressed appropriately. On the other hand, in a periodic component in which periodic vocal cord vibration is a sound source, the shape and timing of the sound source waveform are very important, and it is necessary to accurately represent these in order to realize good sound quality.

  The present invention has been made in view of the above circumstances, and provides a speech processing apparatus, method, and program capable of fusing speech segments without destroying the characteristics of the sound source and vocal tract filter inherent in the speech waveform. The purpose is to do.

  In order to solve the above-described problems and achieve the object, a speech processing apparatus according to an aspect of the present invention includes a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit, and each of the plurality of segments. A phoneme / prosody input receiving unit that accepts input of prosodic information corresponding to, and a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment for each of the segments An acquisition unit, a vocal tract filter component fusion unit that fuses the acquired vocal tract filter components of the plurality of speech units for each segment, and a sound source component of the acquired periodic components of the plurality of speech units. The sound source component fusion unit that expands and contracts based on the shape of the fundamental frequency or the sound source component waveform and fuses for each segment and the fusion fused by the vocal tract filter component fusion unit Using a vocal tract filter characterized by a tract filter component, filtering the fused sound source component fused by the sound source component fusion unit, the plurality of speech segments obtained by the obtaining unit are obtained for each segment. And an element fusion part to be fused.

  Further, in the speech processing method according to another aspect of the present invention, the phoneme / prosody input receiving unit corresponds to each of a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit, and each of the plurality of segments. An input receiving step for receiving input of prosodic information, and an acquiring step for acquiring, for each of the plurality of segments, a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment. A vocal tract filter component merging unit for merging the vocal tract filter components of the plurality of acquired speech segments for each segment; and a plurality of sound source component merging units acquired. The sound source component fusion unit that expands and contracts the sound source component of the periodic component of the speech unit based on the fundamental frequency or the shape of the sound source component waveform and fuses it for each segment. And the unit fusion unit filtering the fused sound source component fused in the sound source component fusion step using a vocal tract filter characterized by the fused vocal tract filter component fused in the vocal tract filter component fusion step By doing so, a unit fusion step of fusing the plurality of speech units acquired in the acquisition step for each segment is included.

  Further, in the speech processing program according to another aspect of the present invention, the phoneme / prosody input reception unit corresponds to each of a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit, and each of the plurality of segments. An input receiving step for receiving input of prosodic information, and an acquiring step for acquiring, for each of the plurality of segments, a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment. A vocal tract filter component merging unit for merging the vocal tract filter components of the plurality of acquired speech segments for each segment; and a plurality of sound source component merging units acquired. A sound source that expands and contracts the sound source component of the periodic component of the speech unit based on the fundamental frequency or the shape of the sound source component waveform and fuses it for each segment A fused sound source component fused in the sound source component fusion step using a vocal tract filter characterized by a fused voice tract filter component fused in the vocal tract filter component fusion step, , By causing the computer to execute a segment fusion step of fusing the plurality of speech segments acquired in the acquisition step for each segment.

  According to the present invention, there is an effect that speech segments can be fused without destroying the characteristics of the sound source and vocal tract filter inherent in the speech waveform.

It is a block diagram which shows an example of a structure of the audio | voice processing apparatus of this Embodiment. It is a figure which shows an example of the information memorize | stored in the speech unit memory | storage part of this Embodiment. It is a block diagram which shows an example of a detailed structure of the element fusion part of this Embodiment. It is a figure which shows an example of the process of the fusion unit extraction part of this Embodiment. It is a block diagram which shows an example of a detailed structure of the vocal tract filter component fusion part of this Embodiment. It is a block diagram which shows an example of a detailed structure of the sound source component fusion | melting part of this Embodiment. It is a figure for demonstrating an example of the deformation | transformation method of the sound source waveform by the sound source waveform deformation | transformation part of this Embodiment. It is a figure for demonstrating an example of the production | generation process of the audio | voice waveform by the production | generation part of this Embodiment. It is a flowchart which shows an example of the flow of the process sequence of the speech synthesis performed with the speech processing unit of this Embodiment. It is a flowchart which shows an example of the flow of the process sequence of the speech unit acquisition process of this Embodiment. 12 is a flowchart illustrating an example of a flow of a procedure for creating a fusion speech unit performed by the speech processing apparatus according to the second modification. It is a block diagram which shows an example of a structure of the audio | voice processing apparatus of the modification 3.

  DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, preferred embodiments of a sound processing device, a method, and a program according to the invention will be described in detail with reference to the accompanying drawings.

  FIG. 1 is a block diagram showing an example of the configuration of the speech processing apparatus 1 according to the present embodiment. As shown in FIG. 1, the speech processing apparatus 1 includes a text input unit 10, a language processing unit 20, a prosody processing unit 30, and a speech synthesis unit 40.

  The text input unit 10 inputs text to be subjected to speech processing.

  The language processing unit 20 performs language analysis such as morphological analysis and syntax analysis of text input from the text input unit 10.

  The prosodic control unit 30 processes accents and intonations from the language analysis result of the language processing unit 20 to generate phoneme sequences and prosodic information.

  The speech synthesizer 40 generates a speech waveform from the phoneme sequence and prosodic information generated by the prosody controller 30. The speech synthesis unit 40 includes a phoneme / prosody input reception unit 41, a speech unit storage unit 42, an acquisition unit 43, a unit fusion unit 44, a vocal tract filter component fusion unit 45, and a sound source component fusion unit. 46, a generation unit 47, and an output unit 48.

  The phonological / prosodic input receiving unit 41 receives a plurality of segments obtained by dividing the phonological sequence corresponding to the target speech by the synthesis unit from the prosody control unit 30 and the input of the prosodic information corresponding to each of the plurality of segments. Specifically, the phoneme / prosody input receiving unit 41 divides the phoneme sequence input from the prosody control unit 30 into segments that are synthesis units, and receives together with the prosody information corresponding to each of the plurality of divided segments.

  The “target speech” refers to a (virtual) speech that is a target when synthesizing speech, that is, an ideal natural speech that realizes the arrangement and prosody of input phonemes. The “phoneme sequence” is, for example, a sequence of phoneme symbols, and the “prosodic information” is, for example, a fundamental frequency, a phoneme duration, power, or the like. The “synthesis unit” is a unit of speech used when generating synthesized speech, and is a combination of phonemes or phonemes divided (for example, semi-phonemes). For example, semitones, phonemes (C, V), diphones (CV, VC, VV), triphones (CVC, VCV), syllables (CV, V), etc. (V is a vowel, C is a consonant), It may be variable length such as a mixture of these.

  The speech unit storage unit 42 stores a plurality of speech units and environmental information associated with each speech unit in association with each other. For example, an HDD (Hard Disk Drive), an optical disc, a memory card, This can be realized by an existing storage medium such as a RAM (Random Access Memory). In the present embodiment, the speech processing apparatus 1 includes the speech unit storage unit 42. However, the speech unit storage unit 42 is provided by an external storage medium or the like (for example, a storage medium that is detachable from the speech processing apparatus 1). In the case of realization, the speech unit storage unit 42 may be omitted.

  The “speech segment” indicates a waveform of a speech signal corresponding to a synthesis unit or a parameter series representing the characteristics thereof. The “environment information” is information indicating the phoneme / prosodic environment of the associated speech unit. For example, the phoneme name of the speech unit, the preceding phoneme, the subsequent phoneme, the subsequent phoneme, the fundamental frequency, and the phoneme duration Length, power, presence of stress, position from the accent core, time from breathing, speaking speed, emotion, etc. In addition to these, information such as cepstrum coefficients at the start and end of speech units that are useful for selecting speech units among the acoustic features of speech units should be included in the “environment information”. Also good.

  FIG. 2 is a diagram illustrating an example of information stored in the speech unit storage unit 42. In the example shown in FIG. 2, the speech segment has a speech waveform when the synthesis unit is a phoneme. These speech segments are obtained by cutting speech waveforms for each phoneme from a large number of speech data labeled for each phoneme according to the label.

  In the example shown in FIG. 2, the environment information includes phonemes (phoneme names) corresponding to speech segments, adjacent phonemes (here, two phonemes each before and after), fundamental frequency, phoneme duration length, and acoustic features. This is a cepstrum coefficient at the beginning and end of a speech unit. The environmental information can be obtained by analyzing and extracting the voice data from which the speech segment is cut out.

  Although FIG. 2 shows an example in which the synthesis unit is a phoneme, the synthesis unit may be a semiphoneme, a diphone, a triphone, a syllable, or a combination or variable length thereof.

  Returning to FIG. 1, the acquisition unit 43 acquires, for each of the plurality of segments divided by the phoneme / prosody input reception unit 41, a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment. . Specifically, the acquisition unit 43 acquires a plurality of speech units associated with segments and prosodic information (environment information) corresponding to the segments from the speech unit storage unit 42.

  At this time, the acquisition unit 43 uses each speech unit candidate as a scale for acquiring speech units, as in the existing unit selection type speech synthesis method and multiple unit selection fusion type speech synthesis method. Using a cost that indirectly represents the magnitude of distortion between the synthesized speech and the target speech in the case of synthesis, a combination of speech units to be merged so as to minimize this cost is acquired.

  Note that the cost used as a measure of speech segment acquisition includes the target cost that represents the degree of distortion of the synthesized speech generated by using the target speech segment in the target phoneme / prosodic environment and the target speech. It consists of a connection cost that represents the degree of distortion of the synthesized speech with respect to the target speech that occurs when the segment is connected to an adjacent speech segment.

  The target cost includes the distortion (basic frequency cost) caused by the difference (difference) between the fundamental frequency of the speech unit and the target fundamental frequency (difference), the difference between the phoneme duration length of the speech unit and the target phoneme duration length (difference) ) Due to the difference between the phoneme environment to which the speech segment belonged and the target phoneme environment (phoneme environment cost), within the word or exhalation paragraph where the speech segment originally existed, There is distortion (positional environmental cost) caused by the difference between the position in the sentence and the position at the time of composition.

  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.

  The unit fusion unit 44 unites a plurality of speech units acquired by the acquisition unit 43 to generate a new speech unit. Specifically, the unit fusion unit 44 uses a vocal tract filter characterized by a fused vocal tract filter component fused by a vocal tract filter component fusion unit 45 described later, and fuses it in a sound source component fusion unit 46 described later. By filtering the fused sound source components, the plurality of speech segments acquired by the acquisition unit 43 are fused for each segment.

  FIG. 3 is a block diagram illustrating an example of a detailed configuration of the segment fusion unit 44 of the present embodiment. As shown in FIG. 3, the unit fusion unit 44 includes a multiple unit input reception unit 441, a fusion unit extraction unit 442, a pre-emphasis unit 443, a linear prediction analysis unit 444, a de-emphasis unit 445, a target A power calculation unit 446, a linear prediction filter unit 447, a power correction unit 448, and a fusion speech unit output unit 449 are included.

  The multi-unit input receiving unit 441 receives input of speech units acquired by the acquiring unit 43 for each segment.

  The fusion unit extraction unit 442 extracts a waveform of a fusion unit suitable for fusion from each of a plurality of speech units input for each segment, and aligns the number of waveforms of each speech unit for each segment.

  In this embodiment, the unit of fusion is a pitch waveform. A “pitch waveform” is a relatively short waveform that has a length that is several times the basic period of speech and does not have a basic period.

  As a method for extracting such a pitch waveform, for example, there is a method using a basic period synchronization window. In this method, a mark (pitch mark) is added to the speech waveform of each speech unit in advance at every basic period interval, and Hanning whose window length is twice the basic period centering on this pitch mark. A pitch waveform is cut out by windowing with a window.

  FIG. 4 is a diagram illustrating an example of processing of the fusion unit extraction unit 442 when the fusion unit is a pitch waveform.

  In the example shown in FIG. 4, three types of speech waveforms surrounded by a dotted line 60 indicate speech waveforms of speech segments acquired for a certain segment. Further, the three types of speech waveforms surrounded by the dotted line 61 indicate pitch waveform sequences extracted from the three types of speech waveforms surrounded by the dotted line 60 by a method using a basic period synchronization window. As shown in FIG. 4, the number of pitch waveforms extracted from the speech waveform usually differs for each speech unit.

  Therefore, the fusion unit extraction unit 442 aligns the number of pitch waveforms of each speech unit to the same number for each segment. Specifically, the fusion unit extraction unit 442 increases the number of pitch waveforms by duplicating several pitch waveforms included in the sequence for sequences with a small pitch waveform, and for sequences with a large pitch waveform. Reduces the number of pitch waveforms by thinning out several pitch waveforms in the sequence.

  In the present embodiment, 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. For example, the number of pitch waveforms is the largest. You may make it align with a thing.

  Then, like the three types of speech waveforms surrounded by the dotted line 62 in FIG. 4, the number of pitch waveforms of each speech unit is made equal to the same number for each segment. Note that three types of speech waveforms surrounded by a dotted line 62 show an example in which the number of pitch waveforms is set to seven.

  Returning to FIG. 3, the pre-emphasis unit 443 applies a negative slope (called tilt, generally seen in the spectral envelope of the sound) to each of the sound waveforms to be fused, from a low frequency to a high frequency. Perform filtering to remove power). Specifically, the pre-emphasis unit 443 applies high-frequency components to each pitch waveform extracted by the fusion unit extraction unit 442 so as to remove the negative slope (tilt) in the entire spectrum envelope. Perform emphasis filtering.

  Here, the voiced sound source is the vibration of the expiratory airflow (voiced volume flow) caused by the periodic opening and closing of the vocal cords, but since the frequency characteristics of this sound source waveform have a low-pass characteristic, the voice waveform of the voiced sound The above-described tilt is generally seen in the frequency component. For this reason, it is preferable to remove such tilt in advance in order to accurately analyze the spectral characteristics of the vocal tract filter.

  Therefore, in the present embodiment, the pre-emphasis unit 443 performs filtering by using a pre-emphasis filter that is generally used in speech analysis, that is, a filter whose transfer function is expressed by, for example, Equation (1). Do. Note that a value of 0.98 to 1.0 is normally used for a in the formula (1). The value of a may be constant for all pitch waveforms, or may be changed for each pitch waveform based on the position in the sentence where the speech segment originally existed. For example, at the end of the sentence, the tension of the vocal cords tends to relax and the tilt tends to increase, so the value of a may be set larger than the rest.

  The linear prediction analysis unit 444 performs linear prediction analysis on each of the speech waveforms filtered by the pre-emphasis unit 443, and calculates a linear prediction coefficient and a linear prediction residual. Here, when the speech waveform to be analyzed is s (n), the linear prediction coefficient is αk (k = 1,..., P is the order of analysis), and the linear prediction residual is e (n), these relationships Is expressed by the following mathematical formula (2).

  In the linear prediction analysis, a linear prediction coefficient is obtained so as to minimize the mean square of the linear prediction residual e (n) in Equation (2).

  Although Equation (2) is an all-pole filter, since the vocal tract system function can be well approximated by an all-pole filter in the speech generation model, this linear prediction is performed in this embodiment. Consider the filter as a vocal tract filter. That is, the linear prediction coefficient obtained by the linear prediction analysis represents the spectral characteristics of the vocal tract filter, and the linear prediction residual is regarded as an approximation of the sound source waveform.

  As a method of linear prediction analysis, an existing method such as an autocorrelation method or a covariance method may be used. In this embodiment, for example, when the original speech waveform is 22 kHz sampling, the analysis order p is set to a value of about 20.

  Then, as described above, the linear prediction analysis unit 444 calculates a linear prediction coefficient and a linear prediction residual for each of a plurality of pitch waveforms for the segment by linear prediction analysis, and uses the linear prediction coefficient as a vocal tract. The result is output to the filter component fusion unit 45, and the linear prediction residual is output to the de-emphasis unit 445.

  The de-emphasis unit 445 performs inverse filtering of the filtering applied by the pre-emphasis unit 443 on each of the linear prediction residual waveforms calculated by the linear prediction analysis unit 444, and uses the de-emphasized linear prediction residual as a sound source component. The data is output to the fusion unit 46.

  That is, the de-emphasis unit 445 performs filtering that restores the high-frequency emphasis by the pre-emphasis unit 443 using a filter whose transfer function is expressed by, for example, Equation (3). Note that the same value as that used in the pre-emphasis unit 443 is used as the value of a.

  The target power calculation unit 446 calculates a target power that is a target power of a new speech unit generated by the fusion based on the power of the speech waveform input from the fusion unit extraction unit 442. Specifically, the target power calculation unit 446 calculates the target power of a new pitch waveform generated by the fusion from the plurality of pitch waveforms of the segments extracted by the fusion unit extraction unit 442. In the present embodiment, the target power calculation unit 446 calculates the power for each of the pitch waveforms and averages them to obtain the target power.

  The linear prediction filter unit 447 filters the fused sound source component fused by the sound source component fusion unit 46 using a vocal tract filter characterized by the fused vocal tract filter component fused by the vocal tract filter component fusion unit 45. To generate a fused speech segment. Specifically, the linear prediction filter unit 447 uses, for each segment, the fused sound source waveform fused by the sound source component fusion unit 46 using the fused linear prediction coefficient fused by the vocal tract filter component fusion unit 45. By filtering, a pitch waveform of the fusion speech unit is generated.

  The linear prediction filter is expressed by Equation (2), and a linear prediction coefficient that has been merged by the vocal tract filter component fusion unit 45 is used for αk. Also, by substituting the fused sound source waveform fused by the sound source component merging unit 46 into e (n) of Equation (2), a pitch waveform of the fused sound source waveform is generated as s (n).

  The power correction unit 448 amplifies or reduces the power of the speech waveform of the fused speech unit generated by the linear prediction filter unit 447 so as to match the target power calculated by the target power calculation unit 446.

  The fused speech unit output unit 449 outputs the fused speech unit for each segment corrected by the power correcting unit 448 so as to match the target power, to the generating unit 47.

  In this embodiment, the unit fusion unit 44 (linear prediction analysis unit 444) uses a linear prediction analysis method as a method of separating a speech unit into a vocal tract filter component and a sound source component. An existing separation method such as a method using an ARX (Auto Regressive with eXogenous input) speech generation model in which the vocal tract filter is approximated by a pole-zero filter may be used.

  In the voice generation process, the periodic components of voiced sound are: (1) vibration of expiratory airflow (vocal volume) generated by the periodic opening and closing of the vocal cords, and (2) adjusting the shape with the tongue, lips, and palate (articulation) ), And (3) produced by being emitted in the lips or nasal passages.

  By the way, in a method using a linear prediction analysis method or an ARX model, (2) is generally approximated by a vocal tract filter, while (1) includes the effect of radiation of (3), that is, (1) (3) is handled as a sound source.

  However, in the present embodiment, the sound source component does not necessarily include the radiation effect, and may be an approximation of only the component (1). Since the radiation characteristic can be well approximated by differentiation, the sound source component approximating only the component (1) can be obtained by integrating the sound source component obtained by the linear prediction analysis method or the method using the ARX model. .

  Returning to FIG. 1, the vocal tract filter component fusion unit 45 merges the vocal tract filter components representing the characteristics of the vocal tract filter of the plurality of speech segments acquired by the acquisition unit 43 for each segment.

  FIG. 5 is a block diagram showing an example of a detailed configuration of the vocal tract filter component fusion unit 45 of the present embodiment. As shown in FIG. 5, the vocal tract filter component fusion unit 45 includes a multiple linear prediction coefficient input reception unit 451, an LSP conversion unit 452, an LSP averaging unit 453, an LPC conversion unit 454, and a fused linear prediction coefficient output. Part 455.

  The multiple linear prediction coefficient input reception unit 451 receives input of linear prediction coefficients calculated for each of the plurality of pitch waveforms for the segment from the linear prediction analysis unit 444 (see FIG. 3).

  The LSP conversion unit 452 converts each of a plurality of linear prediction coefficients (LPC: Linear Prediction Coefficient) received by the multiple linear prediction coefficient input reception unit 451 into a line spectrum pair (LSP). The “line spectrum pair” is a frequency domain parameter that can be mutually converted with the linear prediction coefficient, and can be converted from the linear prediction coefficient by an existing method.

  The LSP averaging unit 453 converts the plurality of line spectrum pairs converted by the LSP conversion unit 452 into i-th coefficients (for example, the line spectrum pair for the 20th-order linear prediction coefficient is greater than 0 and less than π). It is averaged to be composed of coefficients representing 20 angular frequencies.

  The LPC conversion unit 454 converts the line spectrum pair averaged by the LSP averaging unit 453 into a linear prediction coefficient.

  The fused linear prediction coefficient output unit 455 outputs the linear prediction coefficient converted by the LPC conversion unit 454 to the linear prediction filter unit 447 (see FIG. 3) as a fused linear prediction coefficient.

  In the present embodiment, the line spectrum pair is generally excellent in correspondence with the formant frequency, and the average spectrum characteristic common to the plurality of linear prediction coefficients is relatively reduced by averaging in the line spectrum pair region. Since the vocal tract filter component fusion unit 45 uses the above fusion method because it can be obtained satisfactorily, the fusion method of linear prediction coefficients is not limited to this method.

  For example, the vocal tract filter component fusion unit 45 calculates a linear prediction pole from a linear prediction coefficient and then interpolates a plurality of linear prediction poles to obtain an average linear prediction pole, or converts a linear prediction coefficient into an LPC mel cepstrum. For example, a method of averaging in the mel cepstrum region and returning to the linear prediction coefficient may be used.

  Further, the vocal tract filter component fusion unit 45 receives the input of the original pitch waveform (plurality) instead of the plurality of linear prediction coefficients, and performs linear prediction analysis of those pitch waveforms connected in the time direction. Thus, the vocal tract filter components may be fused by obtaining a linear prediction coefficient that averages the characteristics of a plurality of pitch waveforms.

  In this embodiment, the case where the vocal tract filter component fused by the vocal tract filter component fusion unit 45 is a linear prediction coefficient has been described as an example. However, the vocal tract filter component fused by the vocal tract filter component fusion unit 45 is linear. The parameters are not limited to prediction coefficients, and any parameters may be used as long as they represent the characteristics of the vocal tract filter. For example, when an ARX model is used for the vocal tract filter, the vocal tract filter component fusion unit 45 fuses each filter coefficient of the ARX model.

  Returning to FIG. 1, the sound source component merging unit 46 expands and contracts the sound source component of the periodic components of the plurality of speech units acquired by the acquiring unit 43 based on the fundamental frequency or the shape of the sound source component waveform for each segment. To merge. For example, a method such as PSHF (Pitch-scaled harmonic filter) can be used to separate the speech element into a periodic component and an aperiodic component.

  FIG. 6 is a block diagram illustrating an example of a detailed configuration of the sound source component fusion unit 46 of the present embodiment. As shown in FIG. 6, a multiple sound source waveform input receiving unit 461, a defective sound source waveform removing unit 462, a sound source waveform alignment unit 463, a sound source waveform deforming unit 464, a sound source waveform averaging unit 465, and a fused sound source waveform output Part 466.

  The multiple sound source waveform input receiving unit 461 receives, from the de-emphasis unit 445 (see FIG. 3), an input of a linear prediction residual corresponding to each of a plurality of pitch waveforms for a segment as an input of a sound source waveform.

  The defective sound source waveform removing unit 462 checks each of the plurality of sound source waveforms input to the multiple sound source waveform input receiving unit 461 and removes the sound source component corresponding to a predetermined removal condition.

  Note that the “predetermined removal condition” corresponds to, for example, a case where a plurality of pulse-like components considered to be glottal closing points (other than the end of the sentence) are applicable (corresponding to a portion having a problem in utterance such as a rattling voice) ). In addition, for example, the case where the position of the pulse component in the linear prediction residual waveform is greatly deviated from the center position of the linear prediction residual waveform is applicable (because there is a problem in the cutout position of the original pitch waveform). Further, for example, the case where the shape of the linear prediction residual waveform is greatly different from the shape of other waveforms is applicable (because this corresponds to a case where the style of utterance is greatly different).

  The sound source waveform alignment unit 463 sets each of the plurality of sound source waveforms from which the sound source component corresponding to the predetermined removal condition has been removed by the defective sound source waveform removal unit 462 so that the error of the position of the feature point of the sound source component is equal to or less than the threshold value. Align in the time direction so that Specifically, the sound source waveform alignment unit 463 selects each of the plurality of linear prediction residual waveforms input from the defective sound source waveform removal unit 462, and the most important position in the sound source waveform is between the linear prediction residual waveforms. Align in time direction to match.

  In the present embodiment, the most important position in the sound source waveform is considered as the glottal closing point, and the position corresponding to this glottal closing point is aligned in the time direction among the plurality of linear prediction residual waveforms. The “glottal closing point” represents the timing at which the glottal that has been opened in the glottal oscillation cycle closes abruptly. In the linear prediction residual waveform, the local peak within one fundamental period corresponds to that timing. .

  Therefore, the sound source waveform alignment unit 463 obtains the position of the maximum amplitude for each of the plurality of linear prediction residual waveforms, and aligns them in the time direction so that these positions coincide among the plurality of linear prediction residual waveforms. .

  However, some of the linear prediction residual waveforms do not have a clear glottal closure point, i.e., there is no significant peak, and the maximum amplitude position obtained for these does not correspond to the glottal closure point. There may be cases.

  Therefore, more robust alignment is possible when other indices such as cross-correlation with other linear prediction residual waveforms are taken into consideration. For example, the cost function is minimized with the cost function as the value obtained by dividing the square of the position shift of the maximum amplitude between the linear prediction residual waveforms by the cross-correlation between the linear prediction residual waveforms. Alignment should be done.

  In the present embodiment, the method for obtaining the position of the maximum amplitude of the linear prediction residual waveform has been described as a method for obtaining the glottal closing point, but any method that can appropriately extract the glottal closing point, such as a method using wavelet transform. Any method may be used. Further, the alignment method in the time direction between the linear prediction residual waveforms is not limited to the above method, and the gap of the glottal closing point between the linear prediction residual waveforms falls within a desired range. Any method may be used as long as it is a method.

The sound source waveform deforming unit 464 applies deformation such as expansion and contraction in the time direction to each of the plurality of linear prediction residual waveforms aligned by the sound source waveform alignment unit 463.

  FIG. 7 is a diagram for explaining an example of a method for deforming a sound source waveform by the sound source waveform deforming unit 464, and shows an example of a linear prediction residual waveform input to the sound source waveform deforming unit 464.

  In the linear prediction residual waveform shown in FIG. 7, the length of all sections (Dall) corresponds to one pitch waveform which is a unit of fusion, and the basic period of the speech waveform at the position where the pitch waveform originally existed. It is about twice.

  D1 to D4 represent each section when divided into four sections based on the feature points in the linear prediction residual waveform. Specifically, a position including the position with the maximum amplitude, that is, the position corresponding to the glottal closing point and its periphery is immediately before D3 and D3 and a section having a negative amplitude is D2, and a section in front of D2 is D1 and D3. The rear section is D4.

  In the present embodiment, the sound source waveform is deformed by the sound source waveform deforming unit 464 by expanding and contracting in the time direction based on the fundamental frequency of the target sound.

  The sound source waveform should ideally have a length that matches the fundamental frequency of the target speech, but if the fundamental frequency of the speech to which the speech unit originally belonged differs from the fundamental frequency of the target speech, The waveform length is likely to be inappropriate for the target speech.

  Therefore, the sound source waveform deforming unit 464 expands and contracts in the time direction so that the length of the entire section Dall of the sound source waveform is twice the basic period of the target speech (the length obtained by dividing 1 second by the basic frequency). To do.

  However, in the sound source waveform, the shape of the section D3 around the glottal closure point is very important, and if this section is deformed, there is a high possibility that the sound quality will be greatly affected.

  Therefore, in the present embodiment, the sound source waveform deforming unit 464 retains the original shape without deforming the section D3. That is, the sections D1, D2, and D4 are expanded and contracted so that the length of the entire section Dall is twice the basic period of the target speech.

  In the sound source waveform, the shape of the section D2 is also relatively important. However, if the sound source waveform is averaged by the sound source waveform averaging unit 464 while the position of the section D2 is different among a plurality of sound source waveforms, the section The shape of D2 is broken, which can be a factor that degrades the quality of the synthesized sound.

  Therefore, in the expansion / contraction of the section D2 by the sound source waveform deforming unit 464, the expansion / contraction ratio may be determined so that the start points of the section D2 are aligned among a plurality of sound source waveforms.

  In addition, the expansion / contraction method mentioned above is an example, and any expansion / contraction method is applicable.

  The sound source waveform averaging unit 465 averages the plurality of sound source waveforms expanded and contracted by the sound source waveform deforming unit 464 to generate a fused sound source waveform.

  In this embodiment, the linear prediction residual waveform is simply averaged. However, the linearity to be fused using information such as the degree of failure of the sound source waveform output from the defective sound source waveform removing unit 462 or the like. You may weight and average between prediction residual waveforms. Also, after dividing the sound source waveform into multiple frequency bands, after further alignment in the time direction under each frequency band, the sound source waveforms are averaged, and the averaged sound source waveforms of each band are added between the bands. A method of fusing by combining them may be used.

  The fused sound source waveform output unit 466 outputs the fused sound source waveform generated by the sound source waveform averaging unit 465 to the linear prediction filter unit 447 (see FIG. 3).

  In the present embodiment, the sound source component fusion unit 46 performs sound source waveform fusion by averaging the waveforms themselves. However, the sound source waveform is first approximated by a vocal cord sound source wave model, and is used in the parameter region of the model. The sound source waveforms may be merged by averaging and synthesizing the sound source waveform with a vocal cord sound source wave model using the averaged parameters.

  An example of the vocal cord sound source wave model is an LF (Liljencrants and Fant) model. In the LF model, the characteristics of the sound source waveform can be expressed with a high degree of freedom and goodness using five parameters. By approximating each sound source waveform with an LF model and averaging each of these five parameters, it is possible to fuse the sound source waveforms without breaking them.

  When the sound source waveforms are fused in the parameter region of the vocal cord sound wave model, other vocal cord sound wave models such as the Rosenberg model may be used instead of the LF model. Further, as described above, the sound source component fused by the sound source component fusion unit 46 may or may not include an effect due to radiation.

  Returning to FIG. 1, the generation unit 47 deforms and connects the fusion speech units fused by the unit fusion unit 44 to generate a speech waveform of synthesized speech. Specifically, the generation unit 47 changes the fusion speech unit for each segment generated by the unit fusion unit 44 between segments while changing the prosody according to the prosodic information input to the phoneme / prosody input reception unit 41. By connecting, an audio waveform is generated.

  FIG. 8 is a diagram for explaining an example of a voice waveform generation process by the generation unit 47. In FIG. 8, the speech unit “a”, “N”, “s”, “a”, “a” generated by the segment fusion unit 44 is transformed and connected to the speech “aNsaa”. An example of generating a waveform is shown.

  In the example shown in FIG. 8, 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. In addition, the dotted line in FIG. 8 represents the boundary of the segment for each phoneme divided according to the target phoneme duration, and the white triangle indicates the position (pitch mark) where each pitch waveform arranged according to the target fundamental frequency is superimposed. ing.

  As shown in FIG. 8, the generation unit 47 superimposes the pitch waveform of each speech segment on the corresponding pitch mark for voiced sound, and corresponds the waveform of each frame to each frame in the segment for unvoiced sound. By pasting on the part, a speech waveform having a desired prosody (here, fundamental frequency, phoneme duration) is generated.

  The output unit 48 outputs the voice waveform generated by the generation unit 47.

  Next, the operation of the speech processing apparatus 1 according to this embodiment will be described. FIG. 9 is a flowchart showing an example of the flow of a speech synthesis process performed by the speech processing apparatus 1 according to the present embodiment.

  In step S <b> 10, the text input unit 10 inputs text to be subjected to speech processing.

  In step S <b> 11, the language processing unit 20 performs language analysis such as morphological analysis and syntax analysis of the text input from the text input unit 10.

  In step S12, the prosody control unit 30 processes accents and intonations from the language analysis result of the language processing unit 20, and generates phoneme sequences and prosody information.

  In step S13, the phonological / prosodic input receiving unit 41 receives a plurality of segments obtained by dividing the phonological sequence corresponding to the target speech from the prosody control unit 30 in synthesis units and the prosodic information corresponding to each of the plurality of segments. Accept.

  In step S <b> 14, for each of the plurality of segments divided by the phoneme / prosody input receiving unit 41, the acquisition unit 43 converts a plurality of speech units associated with the segment and the prosodic information corresponding to the segment into speech units. A speech segment acquisition process acquired from the storage unit 42 is performed. Details of the speech segment acquisition process will be described later.

  In step S <b> 15, the vocal tract filter component fusion unit 45 fuses the vocal tract filter components of the plurality of speech segments acquired by the acquisition unit 43 for each segment.

  In step S16, the sound source component merging unit 46 expands / contracts the sound source component of the periodic components of the plurality of speech units acquired by the acquiring unit 43 based on the fundamental frequency or the shape of the sound source component waveform, and fuses each segment. To do.

  In step S17, the segment fusion unit 44 filters the fused sound source component fused by the sound source component fusion unit 46 using the fused vocal tract filter fused by the vocal tract filter component fusion unit 45, thereby obtaining the acquisition unit. The plurality of speech segments acquired by 43 are fused for each segment.

  In step S18, the generation unit 47 deforms and connects the fusion speech units fused by the unit fusion unit 44, and generates a speech waveform of synthesized speech.

  In step S <b> 19, the output unit 48 outputs the speech waveform generated by the generation unit 47.

  Next, the speech segment acquisition process in step S14 of FIG. 9 will be described with reference to FIG. FIG. 10 is a flowchart showing an example of the flow of the speech segment acquisition process in step S14 of FIG. FIG. 10 illustrates an example in which M (M ≧ 2) speech segments are selected for each of N (N ≧ 2) segments.

  In step S <b> 101, the acquisition unit 43 selects a speech unit sequence for each segment from the speech unit group stored in the speech unit storage unit 42. Specifically, the acquisition unit 43 uses the target phoneme sequence / prosodic information and the environment information stored in the speech unit storage unit 42 to generate a speech that minimizes the total cost (total cost) as a sequence. An optimum segment sequence that is a sequence of segments is obtained and selected. The search for the optimum unit sequence can be efficiently performed by using dynamic programming (DP).

  In step S102, the acquisition unit 43 assigns an initial value “1” to a counter i that represents a segment number.

  In step S103, the acquisition unit 43 calculates a cost for each speech segment candidate for the segment i. The cost used in this case is the sum of the target cost of the speech unit candidate and the optimal speech unit of the preceding and following segments (speech unit included in the optimal unit sequence) and the connection cost of the speech unit candidate. .

  In step S104, the acquisition unit 43 uses the calculated cost to select the top M speech units with the lowest cost.

  In step S105, the acquisition unit 43 determines whether the counter i is equal to or less than the number of segments N.

  If the number is less than or equal to the number of segments N (Yes in step S105), the process proceeds to step S106. If the number is not less than the number N (No in step S105), the acquisition unit 43 ends the speech segment acquisition process.

  In step S106, the acquisition unit 43 increments the counter i and calculates a cost for each speech element candidate for the segment i.

  As described above, according to the present embodiment, by separating and merging the features of speech units into the features of the sound source and the vocal tract filter, it is possible to perform the fusion without breaking the sound source and the structure of the spectrum. Synthetic speech with higher sound quality than the multiple unit selection and fusion method can be generated.

  In particular, in the present embodiment, the sound source waveform is deformed based on the fundamental frequency of the target sound and the shape of the sound source waveform, thereby removing the controllable differences among the shape differences between the sound source waveforms as much as possible. Therefore, it is possible to perform fusion that keeps the characteristics of the shape of the sound source waveform well and achieve higher sound quality.

  In this embodiment, since sound source waveforms having an inappropriate shape are detected and these sound source waveforms are removed from the fusion target, there is a problem with the original utterance or some failure in the analysis in the middle. Even if a speech unit that has been played is included, the sound quality is unlikely to deteriorate.

(Modification)
It should be noted 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 constituent elements disclosed in the above embodiments. 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.

(Modification 1)
In the above embodiment, the example in which the acquisition unit 43 acquires a common speech unit for the vocal tract filter component and the sound source component has been described. However, the speech unit suitable for each component is separately acquired. Also good.

  In this case, by changing the weighting method between the sub-costs in the calculation of the cost, which is a measure for obtaining the speech element, for each of the vocal tract filter component and the sound source component, the obtaining unit 43 can obtain the speech element suitable for each component. A piece (speech segment that differs between both components) is acquired.

  Specifically, the vocal tract filter component is particularly susceptible to the influence of the phoneme environment such as the front and back phonemes, and it is important for the sound quality of the synthesized sound to smoothly connect the front and back speech segments to the spectrum. Therefore, the phonological environment cost and the spectrum connection cost are set to be heavy.

  On the other hand, sound source components are affected by the position in the expiratory paragraph or sentence (for example, the tension of the vocal cord changes at the beginning and end of the sentence) Because it is particularly susceptible to strong tension (such as high degree of tension), the weight of positional environmental cost and fundamental frequency cost is set to be heavy.

  In this way, since the speech element used for the fusion of each of the vocal tract filter component and the sound source component is acquired by a method suitable for each component, higher sound quality than the above embodiment can be realized. This is particularly effective when there are limited variations of speech segments. In this case, the number of acquired speech segments may differ between the vocal tract filter component and the sound source component.

  Note that the speech segment acquisition method may be a completely different method for each of the vocal tract filter component and the sound source component.

(Modification 2)
In the above embodiment, the example in which the speech processing device 1 generates and outputs a speech waveform from the fused speech unit has been described. However, the speech processing device 1 may be a device that creates a fused speech unit. In this case, the speech synthesis unit 40 of the speech processing device 1 may not include the generation unit 47 and the output unit 48.

  The operation of the sound processing apparatus according to the second modification will be described. FIG. 11 is a flowchart illustrating an example of a procedure for creating a fusion speech unit performed by the speech processing apparatus according to the second modification.

  First, the process from text input to speech unit fusion (step S201 to step S208) is the same as the process from step S10 to step S17 in the flowchart of FIG. In step S201, a large amount of text such as thousands or tens of thousands of sentences is input by the text input unit. Therefore, in step S208, a large number of fused speech segments are generated by the segment fusion unit.

  In step S209, the unit fusion unit 44 determines how many fusion speech units are to be extracted for each unit type of the fusion speech unit from the generated large amount of fusion speech units.

Here, the segment type refers to a category classified by the phoneme environment of the segment. For example, the segment type / a / is a segment corresponding to the phoneme / a /. The number of segments allocated to each segment type is determined according to the appearance frequency 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 /. Let N _i (N _i ≧ 1) be the number of segments to be allocated to the segment type i.

  In step S210, the segment fusion unit 44 substitutes an initial value “1” for the counter i representing the segment type.

In step S211, the unit fusion unit 44 combines the unit type i fused periodic component unit and the united non-periodic component unit into unit type i fused speech units. N _i that have the highest appearance frequency are extracted from N _i .

  In step S212, the segment fusion unit 44 determines whether or not the counter i is equal to or less than the segment type number N (N ≧ 1).

  If it is less than or equal to the number N of segment types (Yes in step S212), the process proceeds to step S213. Finish creating the piece.

In step S213, unit fusion unit 44 increments the counter i, from the fused speech unit of the fused segment type i by unit fusion unit 44, occurrence frequency ones of the upper portions N _i Extract.

  In this way, even in a speech processing device that does not have a speech unit fusion function, as in Modification 3 described later, the fusion speech unit created by the speech processing apparatus in Modification 2 is stored. By doing so, it is possible to perform speech synthesis using speech segments fused without destroying the features of the sound source and vocal tract filter inherent in the speech waveform. Therefore, it is possible to generate synthesized speech with higher sound quality than the conventional multiple unit selection fusion method.

(Modification 3)
In the above-described embodiment, the example in which the speech processing apparatus 1 generates the fusion speech unit has been described. However, in Modification 3, for example, the fusion speech unit created by the speech processing apparatus in Modification 2 or the like is stored in advance. The voice processing apparatus that is being used will be described.

  In the following, differences from the above embodiment will be mainly described. Constituent elements having the same functions as those of the above embodiment are given the same names and symbols as those of the above embodiment, and Description is omitted.

  FIG. 12 is a block diagram illustrating an example of the configuration of the voice processing device 1001 according to the third modification. The synthesizing unit 1040 of the audio processing device 1001 is different from the audio processing device 1 of the above-described embodiment in that the unit fusion unit 44, the vocal tract filter component fusion unit 45, and the sound source component fusion unit 46 are not provided. The synthesizing unit 1040 is different from the speech processing apparatus 1 of the above-described embodiment in that it includes a fused speech unit storage unit 1042 that stores fused speech elements. The acquisition unit 1043 is different from the speech processing apparatus 1 according to the above-described embodiment in that a fusion speech element is obtained from the fusion speech unit storage unit 1042.

  The fused speech unit storage unit 1042 stores a speech unit having a high appearance frequency extracted from the fused speech units generated by the speech processing apparatus according to the second modification.

  Note that the number of speech units to be selected for storage in the fused speech unit storage unit 1042 can be arbitrarily determined by a trade-off between the size of the fused speech unit storage unit 1042 and the sound quality of the synthesized speech. For example, if more speech units are selected and stored, the size of the fused speech unit storage unit 1042 increases, but the quality of the synthesized speech can be improved. For example, if the number of speech units is reduced, the quality of the synthesized speech is sacrificed, but the size of the fused speech unit storage unit 1042 can be reduced.

  As described above, according to the voice processing device 1001 of the third modification, since the voice unit fusion process is not required, it can be applied to low-end middleware having a very low CPU specification.

(Modification 4)
In the above-described embodiment, the method of extracting a segment having a high appearance frequency has been described. However, the segment may be extracted using the feature amount of a segment such as a mel cepstrum calculated at both ends of the segment.

  In this case, the fused periodic component segments and the fused non-periodic component segments output for each segment type are clustered using the segment feature values, and the center (centicent) of each divided cluster is stored. The segment closest to Lloyd) is extracted. 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 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.

  Note that the audio processing apparatuses 1 and 1001 of the above embodiment include a control device such as a CPU (Central Processing Unit), a storage device such as a ROM (Read Only Memory) and a RAM (Random Access Memory), an HDD, an optical disk, and a memory card. The hardware configuration includes an input device such as a touch panel and operation buttons, an audio output device such as a speaker, and the like.

  In addition, the audio processing program executed by the audio processing apparatuses 1 and 1001 of the above embodiment is a file in an installable format or an executable format, and is a CD-ROM, flexible disk (FD), CD-R, DVD ( And recorded on a computer-readable recording medium such as a digital versatile disk.

  In addition, the voice processing program executed by the voice processing apparatus 1 or 1001 of the above embodiment is stored on a computer connected to a network such as the Internet, and is provided by being downloaded via the network. Also good. Further, the voice processing program executed by the voice processing apparatuses 1 and 1001 according to the above-described embodiment may be provided or distributed via a network such as the Internet.

  Further, the voice processing program executed by the voice processing apparatuses 1 and 1001 of the above-described embodiments may be provided by being incorporated in advance in a ROM or the like.

  In addition, the speech processing program executed by the speech processing apparatuses 1 and 1001 of the above embodiment includes the above-described units (phoneme / prosody input reception unit, acquisition unit, segment fusion unit, vocal tract filter component fusion unit, sound source component) The module configuration includes a fusion portion and the like. As actual hardware, a CPU (processor) reads out and executes a translation program from the storage medium, so that the above-described units are loaded onto the main storage device, and a phoneme / prosody input reception unit, acquisition unit, and unit fusion Sections, vocal tract filter component fusion portions, sound source component fusion portions, and the like are generated on the main storage device.

DESCRIPTION OF SYMBOLS 1,1001 Speech processing apparatus 41 Phoneme / prosody input reception part 43,1043 Acquisition part 44 Segment fusion part 45 Vocal tract filter component fusion part 46 Sound source component fusion part

Claims (9)

A plurality of segments obtained by dividing the phoneme sequence corresponding to the target speech by synthesis unit, and a phoneme / prosody input receiving unit that receives input of prosody information corresponding to each of the plurality of segments;
For each of the plurality of segments, an acquisition unit that acquires a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment;
A vocal tract filter component fusion unit that fuses the acquired vocal tract filter components of the plurality of speech segments for each of the segments;
A sound source component fusion unit that expands and contracts the sound source component of the acquired periodic component of the plurality of speech units based on the fundamental frequency or the shape of the sound source component waveform, and fuses each segment;
Acquired by the acquisition unit by filtering the fused sound source component fused by the sound source component fusion unit using a vocal tract filter characterized by the fused vocal tract filter component fused by the vocal tract filter component fusion unit A speech processing apparatus, comprising: a segment fusion unit that fuses the plurality of speech segments that are segmented for each segment.   The sound source component fusion unit aligns the sound source components of the plurality of speech units acquired by the acquisition unit in the time direction so that the error of the position of the feature point of the sound source component is equal to or less than a threshold value, The speech processing apparatus according to claim 1, wherein each segment is fused.   The sound source component merging unit removes the sound source component corresponding to a predetermined removal condition from the sound source components of the plurality of speech units acquired by the acquisition unit, and fuses each segment. The speech processing apparatus according to claim 1 or 2, characterized in that:   The sound source component fusion unit averages the parameters obtained by approximating the sound source components of the plurality of speech units acquired by the acquisition unit with a vocal cord sound wave model, and uses the averaged parameters The sound processing apparatus according to claim 1, wherein sound source components are fused for each segment.   The acquisition unit includes, as a plurality of speech units, a plurality of sound source component fusion speech units used for fusion of the sound source components, and a plurality of the sound source component fusion speech units that are different from the vocal tract filter components. The speech processing apparatus according to claim 1, wherein a plurality of vocal tract filter component fusion speech units used for fusion are acquired.   6. The speech processing apparatus according to claim 5, wherein the number of acquired speech elements for sound source component fusion differs from the number of acquired speech elements for vocal tract filter component fusion. A generating unit that generates a speech waveform by connecting the fusion speech units fused by the unit fusion unit for each segment;
The voice processing apparatus according to claim 1, further comprising an output unit that outputs the voice waveform. A phoneme / prosodic input receiving unit, a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit, and an input receiving step for receiving input of prosodic information corresponding to each of the plurality of segments;
An acquisition unit, for each of a plurality of the segments, an acquisition step of acquiring a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment;
A vocal tract filter component merging unit, wherein the vocal tract filter component fusion step of merging the acquired vocal tract filter components of the plurality of speech segments for each segment;
A sound source component fusion step is performed by expanding and contracting the sound source component of the acquired periodic component of the plurality of speech units based on the fundamental frequency or the shape of the sound source component waveform and fusing each segment.
The unit fusion unit filters the fused sound source component fused in the sound source component fusion step by using a vocal tract filter characterized by the fused vocal tract filter component fused in the vocal tract filter component fusion step. And a segment fusion step of fusing the plurality of speech segments acquired in the acquisition step for each segment. A phoneme / prosodic input receiving unit, a plurality of segments obtained by dividing a phoneme sequence corresponding to a target speech by a synthesis unit, and an input receiving step for receiving input of prosodic information corresponding to each of the plurality of segments;
An acquisition unit, for each of a plurality of the segments, an acquisition step of acquiring a plurality of speech segments associated with the segment and the prosodic information corresponding to the segment;
A vocal tract filter component merging unit, wherein the vocal tract filter component fusion step of merging the acquired vocal tract filter components of the plurality of speech segments for each segment;
A sound source component fusion step is performed by expanding and contracting the sound source component of the acquired periodic component of the plurality of speech units based on the fundamental frequency or the shape of the sound source component waveform and fusing each segment.
The unit fusion unit filters the fused sound source component fused in the sound source component fusion step by using a vocal tract filter characterized by the fused vocal tract filter component fused in the vocal tract filter component fusion step. A speech processing program for causing a computer to execute a unit fusion step of fusing the plurality of speech units acquired in the acquisition step for each segment.

Images