JP5252452B2 - SPECTRUM ANALYZER AND SPECTRUM OPERATION DEVICE - Google Patents

Description

  The present invention relates to speech-related technology, and more particularly to a technology for improving parameterization when statistically processing speech.

  A cepstrum is often used as a speech spectrum expression (parameter). For example, an acoustic model used for speech recognition is often a hidden Markov model (HMM), but a cepstrum is often used as an acoustic parameter for learning. Speech parameterization technology using cepstrum has been well studied, and the software necessary for this has been enhanced. In the cepstrum analysis used in speech recognition or the like, a mel frequency expression obtained by converting a frequency on an auditory frequency scale is often used, and a cepstrum coefficient obtained by the cepstrum analysis is called a mel cepstrum.

  The HMM is used not only for speech recognition but also for speech synthesis. FIG. 1 shows a schematic configuration of a conventional speech synthesis system using an HMM. Referring to FIG. 1, a conventional speech synthesis system 50 using an HMM includes a storage device that stores a phoneme HMM 62 depending on a context, a learning unit 60 for learning the phoneme HMM 62, and input text. 66, a synthesis unit 64 for performing speech synthesis using the phoneme HMM 62 for which learning has been completed.

  The learning unit 60 performs a fundamental frequency extraction process on the speech corpus 70 in which a large number of utterances are stored, and the speech waveform of each phoneme in the speech corpus 70, and outputs a fundamental frequency parameter F0. And a spectrum analysis unit 74 for performing spectrum analysis on the speech waveform of each phoneme in the speech corpus 70 and outputting a spectrum parameter (cepstrum coefficient) representing an envelope of the logarithmic power spectrum of the speech. Further, the learning unit 60 uses the F0 parameter from the fundamental frequency extraction unit 72, the spectrum parameter from the spectrum analysis unit 74, and the phoneme label depending on the context of each phoneme in the speech corpus 70 (hereinafter, this label is referred to as “context-dependent label”). The learning data storage unit 76 for storing the learning data including the probability density function of each context-dependent phoneme model of the phoneme HMM 62 by performing statistical processing on the learning data stored in the learning data storage unit 76 And an HMM model learning unit 78 for calculating the parameters. The context includes the accent type of the phrase including the phoneme, the part of speech of the word including the phoneme, the length of the sentence, the position of the phoneme in the sentence, and the like.

  The synthesizing unit 64 performs text analysis on the input text 66, and is a phoneme label string indicating a phoneme string for the text 66, and a phoneme label string (in accordance with a context in which each phoneme is placed in the text 66 ( The text analysis unit 90 for outputting “context-dependent label sequence”) and the phoneme HMM in the phoneme HMM 62 are concatenated according to the context-dependent label sequence from the text analysis unit 90 and given context-dependent label An acoustic parameter generation unit 92 for estimating an acoustic parameter (F0 and spectral parameter) sequence having the highest likelihood for the sequence from these HMM sequences, and sound source generation according to F0 output from the acoustic parameter generation unit 92 A sound source generation unit 94 for performing sound parameter generation on the sound source waveform from the sound source generation unit 94 By modulating according to the spectral parameter output from the section 92, and a synthesis filter 96 for outputting a synthesized speech signal.

  In such a speech synthesis system 50, it is necessary to learn the phoneme HMM 62 with a large number of speeches. During this learning, after all, the average of the speech spectrum of a specific phoneme context over all samples is calculated. However, when such processing is performed with a cepstrum, a plurality of speech spectra having different positions (frequencies) of spectrum peaks (formants) are averaged in the cepstrum region. In this case, the following problem occurs.

  With reference to FIG. 2, consider two spectra 110 and 112. These have peaks corresponding to formants, but their positions on the frequency axis are shifted from each other. If these are simply averaged, a spectrum 116 is obtained. In the spectrum 116, the peaks clearly existing in the spectrum 110 and the spectrum 112 are lost. If speech synthesis is performed with this spectrum, it is clear that the sound quality is lowered. Originally, as in the spectrum 114, the average of both should be calculated so that a peak clearly occurs.

Omuro Naka et al., “Examination of Inverse Integral Spectrum Function (IFIS) and its Applications”, IEICE Technical Report, SP89-72, p. 23-30, 1989

  One technique for solving such a problem is disclosed in Non-Patent Document 1. Non-Patent Document 1 proposes to use a parameter called “integral spectral inverse function (IFIS)” to interpolate the spectrum. Since the average can be considered as a part of the interpolation, the parameter proposed in Non-Patent Document 1 may be applicable to the above-described processing.

  Referring to FIG. 3, according to this method, when calculating the average of two spectra 130 and 132, the graphs are first integrated over the whole. In the curves 140 and 142 obtained as a result of the integration, frequency values corresponding to the peaks A and B of the original spectra 130 and 132 are obtained (A ′ and B ′), and the average C ′ on these frequency axes is obtained. calculate. This frequency C ′ becomes the center frequency position of the peak C of the spectrum obtained by averaging the spectra 130 and 132. According to Non-Patent Document 1, when the spectra 130 and 132 are further averaged using this technique, the height of the resulting spectrum at each frequency is the height of these spectra 130 and 132 at that frequency. Harmonic average.

  That is, it can be said that the method according to Non-Patent Document 1 is a method of “integrating two spectra from frequency 0 and taking the harmonic average of two points at which the integrated values are equal”.

  An example of a spectrum obtained by averaging the spectra 130 and 132 by this method is shown in FIG. In FIG. 4, the peak of the spectrum 150 is in the middle position between the peaks of the spectra 130 and 132, and the height of the peak is also in the middle of both peaks. Therefore, compared with the example as shown in FIG. It should be noted that the spectra 130 and 132 shown in FIG. 4 are data for testing, and therefore are different from ordinary spectrum curves.

  Certainly, according to the technique according to Non-Patent Document 1, even if two spectra are “averaged”, no peak is lost, which is preferable for HMM learning. However, in the disclosure of Non-Patent Document 1, first, the numerical calculation method of IFIS is not clarified. When the spectrum is actually integrated and its inverse function is taken, there is a problem that requires a large amount of calculation. The obtained IFIS parameters have the same number of dimensions as the spectrum (about 128 to 1024 dimensions). If such high-dimensional acoustic parameters are used for HMM learning, a huge amount of processing (processing time) is required, which becomes a problem. Further, since IFIS is based on simple integration of the spectrum, there is a problem that the frequency resolution is deteriorated in a frequency region where the power is low when the target spectrum has a large slope, for example.

  Therefore, there is a speech parameterization technique that can efficiently calculate a result as shown in FIG. 4 and obtain a parameter suitable for HMM learning and a sufficient frequency resolution over the entire frequency band. is necessary.

  Therefore, an object of the present invention is to provide a spectrum analyzer capable of outputting a parameter that can easily perform shape interpolation between a plurality of spectra without losing characteristic portions of the shapes of the plurality of spectra. Is to provide.

  Another object of the present invention is to provide a spectrum calculation apparatus that can easily perform shape interpolation between a plurality of spectra without losing a characteristic portion of the shape of the plurality of spectra.

  A spectrum analysis apparatus according to a first aspect of the present invention includes a spectrum analysis unit for performing spectrum analysis on a speech signal and outputting a spectrum signal representing a spectrum envelope of the speech, and a spectrum signal output by the spectrum analysis unit. Parameter generation means for outputting a frequency corresponding to each pulse position as a parameter representing the spectral envelope of the audio signal.

  The spectrum analysis means performs spectrum analysis on the input voice signal and outputs a spectrum signal. This spectral signal represents the spectral envelope of the speech. The parameter generation means outputs the frequency corresponding to each pulse position in the pulse density expression of the spectrum signal output by the spectrum analysis means as a parameter representing the spectrum envelope of the audio signal. This output is used as a parameter representing the spectral envelope of the audio signal.

  The spectral envelope of the audio signal is expressed as a parameter in the form of a frequency corresponding to each pulse position in the pulse density representation. Since the characteristics of the spectral envelope are represented by a series of frequency sequences, the correspondence between the characteristic parts can be accurately represented for a plurality of spectra having similar shapes but different frequency positions of the characteristic parts. . As a result, it is possible to provide a spectrum analyzer capable of outputting parameters that can easily perform shape interpolation without losing characteristic portions of the shape for a plurality of spectra.

  Preferably, the parameter generating means inputs a spectrum signal and inputs a frequency corresponding to each pulse position in a pulse density expression obtained based on delta-sigma modulation which performs quantization with a predetermined threshold value. Output as a parameter representing the spectral envelope.

  Spectral signals can be converted to frequency data using delta-sigma modulation, which is often used for time domain signals.

  More preferably, the spectrum signal output by the spectrum analysis means is a cepstrum coefficient sequence representing the spectrum envelope of speech.

  Cepstrum analysis is often used for the analysis of audio signals, and by processing the spectrum information obtained by cepstrum analysis, while using existing means effectively, a new parameter that represents the spectrum envelope as a frequency sequence, The characteristics of the audio signal can be expressed.

  More preferably, the parameter generation means includes a first storage means for storing a zeroth-order cepstrum coefficient among the cepstrum coefficients output from the spectrum analysis means, and a first-order cepstrum coefficient output from the spectrum analysis means. Thereafter, in the pulse density expression of the spectral envelope represented by the cepstrum coefficients up to a predetermined order, the second storage means for storing the frequency corresponding to each pulse position as a frequency sequence, and stored in the first storage means The 0th-order cepstrum coefficient and the frequency sequence stored in the second storage means are output as parameters.

  The zeroth-order cepstrum coefficient represents the average value of the spectrum. By performing pulse density modulation excluding the average value, the average value of the spectrum becomes 0, and the amount of information and the amount of processing when parameterizing can be reduced.

  The parameter generation means includes a first storage means for storing an average value of the spectrum envelope output from the spectrum analysis means, and each pulse in a pulse density expression of the spectrum obtained by subtracting the average value from the spectrum envelope output from the spectrum analysis means. Second storage means for storing the frequency corresponding to the position as a frequency sequence, and the spectrum envelope average value stored in the first storage means and the frequency sequence stored in the second storage means as parameters It may be output.

  Preferably, the spectrum analysis apparatus further includes parameter compression processing means for performing processing for compressing the frequency string data with respect to the frequency string output by the parameter generation means, and the compressed frequency string data is converted into a parameter representing a spectrum envelope. Are output as all or part of

  More preferably, the parameter compression processing unit compresses the frequency sequence output by the parameter generation unit based on a trigonometric series expansion.

  The spectrum analyzing apparatus further includes a spectrum shaping unit that suppresses or removes a global feature including a slope of the spectrum envelope with respect to a spectrum envelope of the sound output from the spectrum analyzing unit, and the spectrum shaping unit The spectral envelope from which the features are suppressed or removed may be input to the parameter generation unit.

  The spectrum shaping means suppresses or removes global features including the slope of the spectrum envelope by subtracting low-order coefficients of the cepstrum from the cepstrum representing the spectrum envelope of the sound output from the spectrum analysis means. Also good.

  A spectrum calculation device according to a second aspect of the present invention receives any one of the spectrum analysis devices described above and the first and second parameters output from the spectrum analysis device with respect to the first and second spectra, respectively. Interpolating means for performing a predetermined interpolation calculation between the first and second parameters.

  The spectrum analyzer outputs, for a plurality of spectra, a frequency corresponding to each pulse position in the pulse density expression as a parameter representing the spectral envelope of the audio signal. This parameter has the property that it can be interpolated without losing the characteristic part of the spectral envelope. The interpolation means performs a predetermined interpolation operation between the first and second parameters obtained from the first and second spectra by the spectrum analyzer. Therefore, interpolation processing can be performed on the first and second spectra without losing the characteristic portion.

  Preferably, the interpolation means includes averaging means for calculating an average of the corresponding parameters among the first and second parameters.

  An average is calculated as an interpolation operation. When calculating the average of a plurality of spectra, an average spectrum can be obtained without losing the characteristic portions of the spectra.

As described above, according to the present invention, the characteristics of the spectrum envelope are represented by a series of frequency sequences, so that the characteristic portions of a plurality of spectra having similar shapes but different frequency positions of the characteristic portions. Can be accurately represented. As a result, it is possible to provide a spectrum analyzer capable of outputting parameters that can easily perform shape interpolation without losing characteristic portions of the shape for a plurality of spectra.

It is a block diagram of the conventional speech synthesis system 50. FIG. It is a spectrum graph which shows a problem when a spectrum is averaged by the conventional method. It is a figure for demonstrating the parameterization method proposed by the nonpatent literature 1. FIG. It is a graph for demonstrating the spectrum averaged by the parameterization method proposed by the nonpatent literature 1. FIG. It is a figure for demonstrating the calculation method of the average of the spectrum by the pulse density modulation (PDM) employ | adopted by embodiment of this invention. 1 is a block diagram of a speech synthesis system 200 according to an embodiment of the present invention. FIG. 7 is a block diagram of the PDM encoder 220 shown in FIG. 6. FIG. 8 is a block diagram of a delta sigma modulation unit 246 shown in FIG. 7. It is a graph for demonstrating the sine series expansion | deployment performed with the compression process part 250 of FIG. It is a figure which contrasts and shows a spectrum and the pulse train obtained from this spectrum by embodiment of this invention. It is a figure which shows the continuous spectrum of the transition part from / r / to / l /. 12 is a graph showing a comparison between a spectrum that is a true average of the continuous spectrum shown in FIG. 11 and a spectrum obtained by averaging the continuous spectrum using the cepstrum and the PDM parameters according to the embodiment of the present invention.

  In the present specification and drawings, the same parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated.

[basic way of thinking]
With reference to FIG. 5, a spectrum parameterization method and a spectrum average calculation method using the parameters according to the embodiment of the present invention will be described. Consider the case of obtaining the average of two spectra 160 and 162. Both peaks are at different positions on the frequency axis, as is apparent from the figure.

  In this embodiment, the envelope of the voice spectrum is first lifted to suppress the characteristic of the opposite spectrum, and then pulse density modulation (PDM) is performed, and the amplitude (or power) of the spectra 160 and 162 is changed to the pulse density. Convert to pulse trains 170 and 172, respectively. By normalizing each spectrum in advance, the number of pulses output for one spectrum is made constant, and the frequency when each pulse is output is stored. For this purpose, it is only necessary to store only the frequency when the pulse is output.

  There is a one-to-one correspondence between the pulses in the pulse trains 170 and 172. A new pulse train 174 is obtained by averaging the frequencies of the corresponding pulse pairs for all pulse pairs. This pulse train 174 is PDM decoded to obtain a new spectrum 180 that averages the spectra 160 and 162.

[Constitution]
FIG. 6 is a block diagram of the speech synthesis system 200 according to the embodiment of the present invention. Referring to FIG. 6, a speech synthesis system 200 using an HMM according to an embodiment of the present invention is a phoneme HMM depending on a context, and is different from the phoneme HMM 62 shown in FIG. Using the storage device that stores the phoneme HMM 212 using the frequency of the pulse train, the learning unit 210 for learning the phoneme HMM 212, and the phoneme HMM 212 that has been learned in accordance with the input text 66. And a synthesizing unit 214.

  The learning unit 210 is connected to receive the output of the spectrum analysis unit 74 of FIG. 1 in addition to the same configuration as the conventional learning unit 60 shown in FIG. 1 further includes a PDM encoder 220 for performing PDM encoding and converting the frequency information of the pulse train representing the spectrum (referred to as “PDM parameters” or “PDM cepstrum”) and outputting the same. Instead of the learning data storage unit 76, the F0 parameter of a certain speech waveform output from the fundamental frequency extraction unit 72, the context-dependent label sequence given to the corresponding speech waveform given from the speech corpus 70, and the PDM encoder 220 Learning data for storing together given PDM parameters as learning data It differs between that it includes a storage unit 222.

  The learning unit 210 includes an HMM model learning unit 78 similar to the learning unit 60 of FIG. Although the HMM model learning unit 78 itself has exactly the same function as that shown in FIG. 1, the phoneme HMM 212 is learned using learning data including PDM parameters instead of a normal cepstrum as an acoustic parameter related to the spectrum. Therefore, the internal parameters of the phoneme HMM 212 are different from the internal parameters of the phoneme HMM 62 shown in FIG. In particular, the phoneme HMM 212 differs from the phoneme HMM 62 of FIG. 1 in that its output is not a cepstrum but a PDM parameter.

  The synthesizing unit 214 has the same configuration as the synthesizing unit 64 shown in FIG. 1, but the acoustic parameters related to the spectrum output from the phoneme HMM 212 are not cepstrum but PDM parameters, and therefore differ from the synthesizing unit 64 in the following points. ing. That is, the synthesis unit 214 decodes the PDM parameters output from the acoustic parameter generation unit 92 in addition to the configuration of the synthesis unit 64 shown in FIG. A decoder 230 is further included. The PDM decoder 230 can be realized with a simple configuration in which a pulse is generated according to the PDM parameter (pulse frequency data) given from the PDM decoder 230 and then the pulse train is passed through a low-pass filter. Here, it is only necessary to regard the spectrum of speech as a speech signal and generate a pulse train in association with the frequency and time.

FIG. 7 is a block diagram of the PDM encoder 220 shown in FIG. Referring to FIG. 7, the PDM encoder 220 shown in FIG. 6 calculates an average value of spectrum parameters output from the spectrum analysis unit 74 and outputs a signal indicating the average value, An average storage circuit 242 for storing an average value signal output from the average value calculation circuit 240, and an average value stored in the average storage circuit 242 from a spectrum parameter output from the spectrum analysis unit 74. A subtracting circuit 244. In fact, among the cepstral coefficients output from the spectrum analyzer 74, since the 0th coefficient C ₀ corresponds to the average value, the average value calculating circuit 240 extracts only the zeroth order coefficients C ₀ This process can be realized by performing the process, and the subtracting circuit 244 performs a process of extracting only the first and subsequent coefficients from the cepstrum coefficients. The average storage circuit 242 stores the 0th-order cepstrum coefficient and outputs it to the learning data storage unit 222 as a part of the PDM parameter.

  Further, the PDM encoder 220 receives the acoustic parameter after subtracting the average value output from the subtraction circuit 244 as input, performs the delta-sigma modulation with the frequency of the spectrum as a time axis, and converts it into a pulse train representing the spectrum. A delta-sigma modulation unit 246 for storing, a pulse train storage unit 248 for storing a frequency of a spectrum when a pulse is output from the delta-sigma modulation unit 246 for each spectrum to be processed, and a pulse train storage unit 248 A compression processing unit 250 for compressing data by performing sine series expansion on the generated pulse train generation position (ie, frequency) information and storing the compressed data in the learning data storage unit 222.

  FIG. 8 is a block diagram of the delta-sigma modulation unit 246 shown in FIG. Referring to FIG. 8, delta sigma modulation section 246 performs primary delta sigma modulation in the present embodiment, and receives integrator 262 that receives the output of subtraction circuit 244, and the output of integrator 262. Includes a quantizer 266 that outputs a +1 pulse when the value exceeds a certain threshold value, and a feedback unit 268 that feeds back the output of the quantizer 266 to the input of the integrator 262. The pulse train storage unit 248 receives the spectrum signal and the output of the quantizer 266, and stores the frequency of the spectrum when the quantizer 266 outputs a pulse. The spectrum waveform is integrated by the integrator 262, and a pulse is output when a certain threshold value is exceeded. Therefore, as shown in FIG. 5, the density of the output pulse is high at the peak portion of the spectrum, and the density of the output pulse is low when the spectrum value is small.

  The compression processing unit 250 shown in FIG. 7 compresses data representing a pulse train (frequency train) by sine series expansion. Referring to FIG. 9, the normalized value of the integral of the spectrum is plotted on the horizontal axis, and the point where each pulse is output is plotted with the frequency axis on the vertical axis, and a curve 290 connecting them is considered. Considering a line segment 292 connecting the first point and the last point in the curve, the curve 290 is divided into an upper portion and a lower portion with the line segment 292 as the center. The example shown in FIG. 9 corresponds to the spectrum 160 shown in FIG. 5, and the curve 290 is divided into two regions: a region existing above the line segment 292 and a region existing below the line segment 292. . Subtracting line segment 292 from curve 290 yields a curve that is just like a sine series. Therefore, the curve 290 can be approximated by a low-order term (about 10 to several tens of degrees) of the sine series expansion, and data indicating each point constituting the curve 290 can be compressed. The same applies to the case where the curve 290 is divided into a plurality of upper and lower portions by the line segment 292. The compression processing unit 250 can represent the frequency string data (hundreds of tens to thousands) with data of a small number of dimensions. As a result, the processing amount in HMM learning is prevented from becoming enormous, and a reasonable calculation is performed. HMM learning can be completed in time.

[Operation]
The operation of the speech synthesis system 200 has two phases. The first phase is learning of the phoneme HMM 212. The second phase is a process of synthesizing speech according to the text 66 using the phoneme HMM 212. Hereinafter, the operation of the speech synthesis system 200 in these phases will be described.

-Learning-
When learning the phoneme HMM 212, the speech synthesis system 200 operates as follows. Each frame of speech in the speech in the speech corpus 70 has a context-dependent label attached in advance. The speech waveform data of each frame is given to the fundamental frequency extraction unit 72 and the spectrum analysis unit 74, respectively. The fundamental frequency extraction unit 72 calculates the F0 parameter from the given voice and provides it to the learning data storage unit 222. The spectrum analysis unit 74 calculates a spectrum parameter of the voice and gives it to the PDM encoder 220.

  Referring to FIG. 7, average value calculation circuit 240 calculates an average value (0th-order cepstrum coefficient) of a given logarithmic power spectrum, and average storage circuit 242 stores the value. The subtracting circuit 244 subtracts the average value from the logarithmic power spectrum, and gives it to the delta-sigma modulation unit 246 starting from the lower frequency.

  Referring to FIG. 8, integrator 262 integrates a given power spectrum. The output of integrator 262 is provided to quantizer 266. The quantizer 266 outputs a pulse when the output of the integrator 262 becomes higher than the threshold value. This pulse is fed back to the input of the integrator 262 by the feedback unit 264. As a result, a value corresponding to the threshold value is subtracted from the output of the integrator 262. In this way, the delta-sigma modulation unit 246 integrates the waveform from the low frequency region to the high frequency region of the spectrum waveform, and outputs a pulse when the integrated value exceeds the threshold value. When a pulse is output, the value corresponding to the threshold value is subtracted from the integrated value. As a result, when the spectrum waveform is integrated from the low frequency side, the integrated value becomes the threshold value. A pulse is output at the time. The pulse train storage unit 248 stores the frequency of the spectrum at this time. Thus, the pulse train storage unit 248 stores the pulse train output for the spectrum in the form of a frequency train when they are output. The compression processing unit 250 converts the frequency sequence data (on the order of hundreds to tenths) stored in the pulse sequence storage unit 248 into decimal-order data by sine series expansion. As a result of this compression processing, it is possible to prevent the amount of processing in HMM learning from becoming enormous and to complete HMM learning in a reasonable calculation time.

  The PDM encoder 220 supplies the frequency sequence data stored in the pulse sequence storage unit 248 and compressed by the compression processing unit 250 to the learning data storage unit 222 as a PDM parameter. The learning data storage unit 222 receives the F0 parameter from the fundamental frequency extraction unit 72 and the context-dependent label from the speech corpus 70 for each frame, and further receives the PDM parameters from the PDM encoder 220 and collects them together. Save as learning data.

  In this way, when learning data is created for all frames of speech data stored in the speech corpus 70, the HMM model learning unit 78 uses this learning data to learn the phoneme HMM 212.

  FIG. 10 shows an example of a logarithmic power spectrum graph of a certain voice and a pulse train obtained when PDM is performed on this spectrum. As shown in FIG. 10, when the logarithmic power of the spectrum is large, the pulse density is high, and when it is small, the pulse density is low.

-Speech synthesis-
At the time of speech synthesis, the speech synthesis system 200 operates as follows. When the text 66 to be synthesized is given, the text analysis unit 90 of the synthesis unit 214 performs a known text analysis process, generates a context-dependent label sequence including a phoneme sequence to be synthesized, and generates an acoustic parameter generation unit. 92.

  The acoustic parameter generation unit 92 concatenates the context-dependent HMMs in the phoneme HMM 212 according to the given context-dependent label sequence. The acoustic parameter generation unit 92 further generates an acoustic parameter string (F0 parameter string and PDM parameter string) having the highest likelihood based on the context-dependent HMM after connection. At this time, since the phoneme HMM 212 performs learning using the PDM parameter as the acoustic parameter, the acoustic parameter sequence output from the acoustic parameter generation unit 92 is not a cepstrum sequence as shown in the related art, but a PDM parameter sequence. It becomes.

  The PDM decoder 230 performs decoding on the PDM parameter sequence output from the acoustic parameter generation unit 92, converts it into a spectral parameter sequence, and provides the resultant to the synthesis filter 96.

  The sound source generation unit 94 generates a sound source signal according to the F0 parameter sequence given from the acoustic parameter generation unit 92, and the synthesis filter 96 performs filtering processing of characteristics depending on the spectral parameter given from the PDM decoder 230 to the sound source signal. To output a synthesized speech signal. By synthesizing and amplifying the synthesized voice signal and giving it to the speaker, the synthesized voice is uttered.

  As described above, according to the present embodiment, spectrum flattening can be mitigated in the spectrum averaging process performed during HMM model learning. Therefore, an effect that the sound quality of the synthesized speech is improved can be obtained.

  For example, consider the case of averaging the continuous spectrum of the transition from / r / to / l / as shown in FIG. Referring to FIG. 12, when a waveform 322 obtained by averaging spectra in the cepstrum region as in the prior art is compared with a waveform 320 averaged with a strict correspondence of frequencies, the waveform 320 has a second formant and a second waveform. While the three formants can be distinguished, the waveform 322 is flattened and cannot be distinguished. On the other hand, in the waveform 324 obtained according to the method of the present embodiment, these formants are properly distinguished, and the characteristics of the waveform calculated strictly can be well preserved.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are included. Including.

50, 200 Speech synthesis system 60, 210 Learning unit 62, 212 Context-dependent phoneme HMM
64, 214 Synthesis unit 66 Text 70 Speech corpus 72 Fundamental frequency extraction unit 74 Spectrum analysis unit 76, 222 Learning data storage unit 78 HMM model learning unit 90 Text analysis unit 92 Sound parameter generation unit 94 Sound source generation unit 96 Synthesis filters 170, 172 , 174 Pulse train 220 PDM encoder 230 PDM decoder 246 Delta-sigma modulation section

Claims (11)

A spectrum analyzer using a pulse density expression that expresses a distribution based on intensity on a frequency axis of a spectrum signal by converting it into a density of pulses generated at frequency positions at variable intervals on the frequency axis,
Spectrum analysis means for performing spectrum analysis on the speech signal and outputting a spectrum signal representing the spectrum envelope of the speech;
In the pulse density representation of the output spectrum signal by said spectrum analysis means, the frequency corresponding to each pulse position, and a parameter generating means for outputting a parameter representing the spectrum envelope of the speech signal, the spectrum analyzer . The frequency corresponding to each pulse position in the pulse density expression obtained based on delta-sigma modulation with the spectrum signal as input and quantizing with a predetermined threshold is used as a parameter representing the spectrum envelope of the audio signal. The spectrum analysis apparatus according to claim 1, wherein the parameter generation means outputs. The spectrum analysis apparatus according to claim 1 or 2, wherein the spectrum signal output by the spectrum analysis means is a cepstrum coefficient sequence representing a spectrum envelope of speech. The parameter generation means includes
Of the cepstrum coefficients output by the spectrum analyzing means, first storage means for storing a zeroth-order cepstrum coefficient;
A frequency corresponding to each pulse position in the pulse density representation of the spectral envelope represented by the cepstrum coefficients from the first order to the predetermined order among the cepstrum coefficients output from the spectrum analyzing means is stored as a frequency sequence. Storage means,
The said 0th-order cepstrum coefficient memorize | stored in the said 1st memory | storage means and the frequency sequence memorize | stored in the said 2nd memory | storage means are output as the said parameter, The Claim 3 characterized by the above-mentioned. Spectrum analyzer. The parameter generation means includes
First storage means for storing an average value of a spectrum envelope output from the spectrum analysis means;
A second storage means for storing a frequency corresponding to each pulse position as a frequency sequence in a pulse density representation of a spectrum obtained by subtracting an average value from a spectrum envelope output from the spectrum analysis means;
3. The spectrum envelope average value stored in the first storage unit and the frequency sequence stored in the second storage unit are output as the parameters. The spectrum analyzer described. Parameter compression processing means for compressing frequency string data for the frequency string output from the parameter generation means, and the compressed frequency string data is converted to all or a part of parameters representing the spectrum envelope. The spectrum analyzer according to any one of claims 1 to 5, wherein The spectrum analysis apparatus according to claim 6, wherein the parameter compression processing unit compresses the frequency sequence output from the parameter generation unit based on a trigonometric series expansion. Spectral shaping means that suppresses or eliminates global features including the slope of the spectral envelope with respect to the spectral envelope of the speech output by the spectral analysis means, and the global features are suppressed or removed in the spectral shaping means. The spectrum analysis apparatus according to claim 1, wherein the removed spectrum envelope is input to the parameter generation unit. The spectrum shaping unit suppresses a global feature including the slope of the spectrum envelope by subtracting a low-order coefficient of the cepstrum from the cepstrum representing the spectrum envelope of the sound output from the spectrum analysis unit. The spectrum analyzer according to claim 8, wherein the spectrum analyzer is removed. The spectrum analyzer according to any one of claims 1 to 9,
Interpolation means for receiving first and second parameters output from the spectrum analyzer for the first and second spectra, respectively, and performing a predetermined interpolation operation between the first and second parameters; Including a spectrum calculation device. The spectrum calculation device according to claim 10, wherein the interpolation means includes an averaging means for calculating an average of corresponding parameters among the first and second parameters.