JPH08305396A - Device and method for expanding voice band - Google Patents

Abstract

PURPOSE: To generate a wide band synthetic voice having high articulation using an extracted spectrum envelope and a sound source having a wide band pulse train from a narrow band voice signal. CONSTITUTION: An input signal is stored in a buffer 101 for a fixed time, feature quantity relating to a spectrum envelope is extracted for a signal train stored in the buffer by an analyzing section 102. By using this feature quantity, a sound source pulse train generating section 13 estimates a sound source pulse train for a signal of the buffer 101, and restores an omitted band using an estimated pitch period. An output voice synthesizing section 14 synthesizes voice by weighing frequency in which voice is omitted from the sound source pulse train and the spectrum envelope. Thereby, a voice band can be expanded for a voice signal in which a band is omitted with comparatively simple constitution.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus and method for expanding a band from a narrow band voice signal to a wide band voice signal in an environment where only a narrow band voice signal is obtained.

[0002]

2. Description of the Related Art In recent years, with the development of digital communication networks, the digitization of voice signals has been rapidly progressing, and many techniques have been proposed for processing voice signals using a digital signal processing technique.

A conventional voice band expanding device and voice band expanding method will be described below. In the conventional voice band expanding method, for example, the communication technique SP93-61 (19
93-08), a method using an analysis-synthesis technology and a vector quantization method has been proposed.

FIG. 5 is a block diagram of an apparatus to which a conventional voice band expanding method can be applied. In FIG. 5, 201 is an LPC
An analysis unit, 202 is a vector quantization unit, 203 is a decoding unit, 204 is a narrowband codebook, 205 is a wideband codebook, 206 is a low band restoration unit, 207 is a first high band restoration unit, and 208 is a second band. Is a high-frequency restoration unit, 209 is an addition unit, and 210 is an upsampling unit.

The operation of the speech band expanding apparatus configured as described above will be described below.

First, the input narrow band speech signal is the LPC.
The analysis unit 201 performs linear prediction analysis, and separates the spectrum envelope information, power information, and pitch information. The extracted spectrum envelope information is vector-quantized by the vector quantization unit 202 with reference to the narrowband codebook 204. Based on the output of the vector quantization unit 202, the decoding unit 203 outputs the wideband codebook 205.
To estimate the broadband spectral envelope information. L
From the power information and pitch information obtained by the PC analysis unit 201 and the wideband spectrum envelope information obtained by the decoding unit 203, the low frequency band restoration unit 206 outputs 300H.
A low frequency component equal to or lower than z is generated. Further, the first high-frequency restoration unit 207 generates a high-frequency component of 3400 Hz or higher from the power information and pitch information obtained by the LPC analysis unit 201 and the spectrum envelope information obtained by the decoding unit 203. . The adder 2 adds the output of the up-sampling unit 210 that up-samples the narrowband audio signal and the outputs of the low-frequency restoring unit 206 and the high-frequency restoring unit 207.
The addition is performed at 09 to obtain a wideband audio signal.

Further, in FIG. 5, the first high-frequency restoration unit 20
A voice band expansion method using a second high-frequency restoration unit 208 instead of No. 7 has also been proposed. In the second high-frequency restoration unit 208, based on the power information and pitch information obtained by the LPC analysis unit 201 and the output of the vector quantization unit 202, a high-frequency wave of 3400 Hz or higher is used by using the waveform segment method. Generate a domain component. The rest of the configuration is the same as the method using the first high-frequency restoration unit 207, and the output of the up-sampling unit 210 that up-samples the narrow band audio signal and the output of the low-frequency restoration unit 206 and the high-frequency restoration unit 208. And adder 20
The addition is made in 9 to obtain a wideband audio signal.

[0008]

However, the above-mentioned conventional configuration has a problem that a large number of codebooks are required to perform expansion of the spectrum envelope information with high accuracy. Furthermore, since the above-mentioned conventional configuration is a method of extracting pitch information from a narrow band speech signal and directly creating a synthesized sound source using it as a parameter, there is a problem that sound quality deterioration due to pitch estimation error occurs. The method using the waveform segment method has a problem that it is necessary to have a codebook of segment data separately.

The present invention solves the above-mentioned conventional problems. The expansion of the spectrum envelope information is widened by using a linear mapping function or the like to obtain a wide band spectrum envelope. By adding a pulse train that interpolates the missing band, further widening the band using the distortion component generated by emphasizing the pulse, and combining by using the wideband spectrum envelope and the synthetic sound source. The purpose is to expand a wide voice band.

[0010]

In order to solve the above-mentioned problems, a voice band expanding apparatus of the present invention comprises a buffer for storing a fixed amount of input signals, and a linear prediction for a signal sequence stored in the buffer. An analysis unit that calculates at least a linear prediction coefficient of the coefficient and the PARCOR coefficient; an impulse response calculation unit that calculates an impulse response from the linear prediction coefficient obtained by the analysis unit; and a buffer of the buffer using the linear prediction coefficient as a parameter. An auditory weighting filter section for weighting the difference signal sequence between the output and the output of the first speech synthesizer to simulate human auditory characteristics, and a pitch period is estimated from the output signal of the auditory weighting filter section. With reference to the pitch period estimation unit and the output value from the pitch period estimation unit, the output signal from the auditory weighting filter and the impulse response A pulse generator that inputs the output of the calculator to generate a pulse train, a pulse train that is output from the pulse generator, and one of the linear prediction coefficient and the PARCOR coefficient, and synthesizes an audio signal. A first voice synthesis unit, a frequency weighting filter that performs frequency weighting on the output of the pulse generation unit, an output value of the frequency weighting filter, and either the linear prediction coefficient or the PARCOR coefficient. A second voice synthesizer for inputting and synthesizing voice.

Further, the voice band expanding apparatus of the present invention stores at least a linear prediction coefficient of a linear prediction coefficient and a PARCOR coefficient for a signal sequence stored in the buffer and a buffer for storing a fixed amount of the input signal. An analysis unit for calculating, an impulse response calculation unit for calculating an impulse response from the linear prediction coefficient obtained by the analysis unit, and a linear prediction coefficient or PARCOR coefficient whose band is expanded based on the linear prediction coefficient or PARCOR coefficient. An envelopment enlarging section for estimating the difference, a perceptual weighting filter section for weighting a difference signal sequence between the buffer and the first speech synthesis section using the linear prediction coefficient as a parameter to simulate human auditory characteristics, A pitch period estimation unit that estimates the pitch period from the output signal of the auditory weighting filter unit; With reference to the force value, a pulse generator that generates a pulse train by inputting the output signal from the auditory weighting filter unit and the output of the impulse response calculator, and a pulse train output from the pulse generator, The first speech synthesizer for synthesizing an audio signal by inputting one of the linear prediction coefficient and the PARCOR coefficient; a frequency weighting filter for performing frequency weighting on the output of the pulse generator; A second voice synthesis unit for synthesizing voice by inputting an output value of the frequency weighting filter and an output value from the envelope enlarging unit is provided.

Here, the above-mentioned pulse generator has a frequency weighting filter for performing frequency weighting on the output signal from the auditory weighting filter, and generates a pulse train for restoring the missing band. You may make it have a function.

Further, the pulse generating section may have a function of emphasizing the pulse by increasing the amplitude of the pulse at the pitch cycle interval, using the detected value of the pitch cycle estimating section as the characteristic amount.

Further, the pulse generating section may have a function of outputting a pulse set to the threshold value by using a threshold value with the detected value of the pitch period estimating section as a characteristic amount.

The envelope enlarging unit has a function of inputting the feature amount obtained by the analyzing unit and broadening the feature amount for the spectrum envelope having the feature of the wide band spectrum envelope by using the linear mapping function. May be.

Further, in the voice band expanding method of the present invention, a certain amount of an input signal is stored in a buffer, a feature quantity relating to a spectrum envelope is extracted from a signal sequence stored in the buffer, and the extracted feature is extracted. For the source pulse train estimated from the amount and the signal train stored in the buffer,
A wide-band source pulse train is generated from a priori information on the frequency characteristic of the signal and the estimated pitch information, and a voice is synthesized using the feature amount related to the spectral envelope and the wide-band source pulse train.

Further, in the voice band expanding method of the present invention, a certain amount of input signal is stored in a buffer, and the feature quantity related to the spectrum envelope is extracted from the signal sequence stored in the buffer to express the feature quantity. The missing envelope information in the envelope information is interpolated, and with respect to the sound source pulse train estimated from the feature quantity and the signal train stored in the buffer, it was estimated as prior information about the frequency characteristic of the signal. Generates a wide-band source pulse train from pitch information,
The speech is synthesized using the feature quantity related to the spectrum envelope and the wide-band source pulse train.

[0018]

According to the present invention, the input signal is stored in the buffer for a certain period of time according to the above-mentioned structure, and the PARCOR analysis unit extracts the feature quantities such as the PARCOR coefficient and the linear prediction coefficient related to the spectral envelope from the signal in the buffer. In addition, the synthesized sound source is represented by a plurality of pulse trains from the signal in the buffer and the feature amount related to the spectrum envelope. When the pulse train is determined and generated, a band that does not exist in the input signal is generated by periodically emphasizing the pitch period estimated in advance as a feature amount or adding distortion to the pitch period. Or using the distortion of clipping. With the pulse train generated in this way as the sound source,
Speech synthesis is performed using the PARCOR coefficient or the linear prediction coefficient output from the PARCOR analysis unit. Also,
The spectral envelope information can be broadened using a linear mapping function or a vector quantization method, and speech synthesis can be performed using the sound source. The voice synthesized in this manner is a voice that retains the fine structure of the spectrum and is broadened so as to interpolate the missing band, thereby providing a high-quality and clear synthesized voice.

[0019]

【Example】

(Embodiment 1) Hereinafter, a first embodiment of the present invention will be described. FIG. 1 is a block diagram showing the overall configuration of an apparatus to which the voice band expanding method according to the first embodiment of the present invention can be applied.

In FIG. 1, reference numeral 101 is a buffer which stores an input signal for a certain period of time. 102 is the buffer 10
An analysis unit that extracts a feature amount such as a linear prediction coefficient or a PARCOR coefficient related to the spectral envelope with respect to the input signal sequence stored in 1 is an impulse response calculation unit that calculates an impulse response from the linear prediction coefficient.

To expand the band of the sound source information, a perceptual weighting filter unit for perceptually weighting humans on the frequency axis by subtracting the output signal from the first speech synthesis unit 105 from the signal sequence of the buffer 101. 104, a pitch period estimation unit 106 that estimates a pitch period from the output of the auditory weighting filter unit 104, and a pulse position and a gain based on the feature amount extracted by the analysis unit 102 to generate a pulse train. And a first speech synthesis unit 105 that synthesizes a synthetic sound from the pulse train generated by the pulse generation unit 107 and the linear prediction coefficient.

The pulse generator 107 determines the pulse position and gain so that the error with respect to the input signal from the perceptual weighting filter 104 is minimized, and the pitch period obtained by the pitch period estimator 106 is also calculated. Distortion such as low-frequency emphasis is added to the pulse trains generated in and. Examples of distortion in the low-frequency emphasis include half-wave rectification, full-wave rectification, clipping,
Examples include power processing.

The final output synthesized voice is the pulse generator 10
Using the pulse train determined in step 7 as the input sound source, the emphasis filter unit 108 for the purpose of emphasizing and interpolating the frequency-missing portion, and the linear prediction coefficient obtained by the analysis unit 102 as the feature amount And a second voice synthesis unit 109 for synthesizing

The first embodiment of the present invention described above will be described in detail below with reference to the block diagram of FIG.

First, in the buffer 101, the audio signal is discretely taken in for a fixed time, and this fixed time interval is 240 if the sampling frequency is 16 kHz, for example.
This time unit will be referred to as a "frame" hereinafter. The analysis unit 102 executes the process of calculating the parameter related to the spectrum envelope of the voice for each frame. The signal captured in the buffer 101 is analyzed by the analysis unit 10
At 2, when the observed signal at time i is y (i) and the mth-order autocorrelation value is r (m), the mth-order autocorrelation value is calculated by (Equation 1).

[0026]

[Equation 1]

Based on the autocorrelation value calculated by (Equation 1), the analysis unit 102 analyzes. Analysis unit 10
In 2, the PARCOR coefficient or the linear prediction coefficient is recursively calculated from the autocorrelation value. This calculation method can be easily realized by using a known technique, and is described in, for example, Acoustic and Speech Engineering, Furui Modern Science Co., Ltd. pp.131-136. If the PARCOR coefficient is obtained, the linear prediction coefficient is uniquely obtained. If the linear prediction coefficient is obtained, the PA is uniquely obtained.
The RCOR coefficient is obtained. At this time, the order of analysis is 1
It is calculated as a value of 0th to 25th order. The impulse response calculation unit 103 calculates the impulse response in the system of the linear prediction coefficient value based on the linear prediction coefficient calculated by the analysis unit 102 by applying an impulse as an input.

The calculation of the sound source starts by subtracting the signal synthesized by the first speech synthesizer 105 from the speech signal in the buffer 101. The subtracted signal is
It has, so to speak, the same effect as removing the influence of the previous analysis frame, and becomes the input signal of the perceptual weighting filter unit 104 in the analysis frame. Auditory weighting filter unit 1
04 is configured by a digital filter having the characteristic of (Equation 2).

[0029]

[Equation 2]

In (Equation 2), W (z) is a function of z
Where ai is the i-th order linear prediction coefficient, c^kIs the desired filter feature
A constant that is given to realize the property (for example, about 0.8
Degree) and z are complex variables generated by z transformation. Heavy hearing
The finding filter unit 104 reduces the influence of the quantization error to power.
By adding around the large formant frequency,
Has the effect of providing a synthetic sound with less sense of roughness.
One. Note that in (Equation 2), c ^kIs, for example, 0.8
At this time, the effect as a filter becomes large. Pitch period
The estimation unit 106 uses the long-term prediction filter to determine the pitch cycle.
Estimate the period. As the pitch period estimation unit 106,
Calculate pitch period by calculating correlation value for prediction
Method, and the pitch period is estimated by adding the waveforms.
Method. For example, the least squares prediction error method (M
A method using the SPE method) and the like can be mentioned.

Specifically, in the calculation formula of (Equation 3),
The pitch period is calculated by setting m that gives the maximum value to match (m) as the pitch period.

[0032]

(Equation 3)

In the pulse generator 107, the output sequence from the perceptual weighting filter unit 104 is input, and (Equation 4)
The gain gk (mk) at the position m _k is calculated by

[0034]

[Equation 4]

In (Equation 4), h _i is the impulse response at points separated by i points. Then, ε _{k of} (Equation 5)
The number of pulses set up in advance is determined so that the point m _k that minimizes is the pulse position.

[0036]

(Equation 5)

The number of pulses can be set arbitrarily, but for example, it is preferable to set a pulse of about 8 points for 80 observation points when the input sequence is 16 kHz sampling.

When the low frequency component is missing in the input speech, the estimated pitch period is used for the pulse train determined by (Equation 5) using the estimated value of the pitch period estimation unit 106. By emphasizing the pulse, the fundamental frequency component is restored as the output of the pulse generator 107.

As a method of emphasizing the pulse in the pulse generation section 107, a simple method is to use the estimated value of the pitch period estimation section 106 and calculate one frame from the ratio of the frame length and the output value from the pitch period estimation section 106. The number of pitch periods with respect to the length is calculated, and one frame is multiplied by a constant number (> = 1) of the maximum amplitude pulses from the top to obtain an output pulse train.

An embodiment of pitch enhancement will be described with reference to FIG. In the figure, a is a schematic representation of the estimated pitch cycle interval estimated by the pitch cycle estimation unit 106,
b represents the estimated pulse train in the current estimated frame calculated by (Equation 5) and the pulse train determined at the previous analysis time, with the vertical axis representing the gain and the horizontal axis representing the time. c is the estimated pulse train after the emphasis processing. b
In the above, among the emphasized pulses in the past estimated frame, it is observed whether or not there is a pulse determined in (Equation 5) at each estimated pitch interval indicated by a from the pulse emphasized at the last time, If there is a corresponding pulse,
The pulse is emphasized as shown in FIG.

Besides, using the estimated value of the pitch period estimation unit 106, the number of pitch periods for one frame length is calculated from the ratio of the frame length and the output value from the pitch period estimation unit 106. However, for one frame, a set of pulses at a position corresponding to the pitch period is searched for, and the number of pitches that should be in one frame, which is obtained from the ratio of the frame length and the output value of the pitch period estimation unit 106, from the set having a large amplitude. However, the pulse is multiplied by a constant (> = 1) to obtain an output pulse train. The first speech synthesis unit 105 receives the output pulse train of the pulse generation unit 107 and outputs the PARCOR output from the analysis unit 102.
A synthetic sound is obtained by PARCOR synthesis using the coefficient as a feature amount.

The output voice is synthesized by using the pulse train generated by the pulse generator 107 as an input and the emphasis filter 10
8, the missing band is restored by an FIR filter or the like. In the second speech synthesis unit 109, the emphasis filter unit 10
The output from 8 is input, and the PARCO from the analysis unit 102 is input.
A synthesized sound obtained by PARCOR synthesis is output using the R coefficient as a feature amount. In this way, the voice is synthesized for each frame.

As described above, according to the voice band expanding apparatus of this embodiment, the buffer 1 for storing the input signal for a certain period of time.
01, and an analysis unit 102 that extracts a feature amount related to the spectrum envelope from the signal sequence stored in the buffer.
And a pulse generator 107 for estimating a sound source pulse for the signal of the buffer 101 using the feature amount from the analyzer 102, and emphasizing a pulse corresponding to the pitch period using the estimated pitch period; By providing the weighting enhancement filter unit 108, it is possible to provide a voice band expanding device that can expand the voice band from a voice signal having a missing band by the effect of pitch enhancement with a relatively simple configuration. .

In this embodiment, the first voice synthesizer 10
5 and the second speech synthesis unit 109, the PARCOR coefficient output from the analysis unit 102 as a feature amount
Synthesized sound is obtained by COR synthesis, but PARCOR
The speech synthesis may be performed using a linear prediction coefficient instead of the coefficient.

(Embodiment 2) Next, a second embodiment of the present invention will be described.

The overall configuration of the second embodiment of the present invention is the same as that of the first embodiment (FIG. 1), but the pulse generator 1 of FIG.
The method of pulse generation in 07 is different. Hereinafter, only this pulse generation method will be described, and other description will be omitted.

An estimated pulse train is obtained from the output from the perceptual weighting filter unit 104 and the output from the impulse response calculation unit 103 by the same method as in the first embodiment, and then the estimated value of the pitch period estimation unit 106 is used. Calculate how many pitch periods there are for one frame length,
From the top, the number of maximum amplitude pulses is set to a certain threshold value, and the estimated pulse above the threshold value is set to a pulse having the magnitude of the threshold value, so that only the output pulse train is different from the first embodiment. .

As described above, according to the voice band expanding apparatus of this embodiment, the buffer 1 for storing the input signal for a certain period of time.
01, and an analysis unit 102 that extracts a feature amount related to the spectrum envelope from the signal sequence stored in the buffer.
And estimating a sound source pulse for the signal of the buffer 101 using the feature amount from the analysis unit 102, and using the estimated pitch period to obtain a pulse corresponding to the pitch period,
By using a certain threshold value and aligning a pulse having an amplitude equal to or larger than the threshold value with the threshold value, distortion is caused in the sound source, and the distortion causes the pulse train 107 to generate a pulse train effective for low-frequency emphasis. By providing the weighting enhancement filter unit 108, a voice band expanding device capable of expanding a voice band from a voice signal with a missing band by using the effect of distortion with respect to the pitch period with a relatively simple configuration. Can be provided.

In this embodiment, the first voice synthesizer 10
5 and the second speech synthesis unit 109, the PARCOR coefficient output from the analysis unit 102 as a feature amount
Synthesized sound is obtained by COR synthesis, but PARCOR
The speech synthesis may be performed using a linear prediction coefficient instead of the coefficient.

(Embodiment 3) Next, a third embodiment of the present invention will be described.

The overall configuration of the third embodiment of the present invention is shown in FIG. The difference from the first embodiment is that the envelope expanding section 110 is provided as shown in FIG.
However, the point is that a linear prediction coefficient is calculated as the feature amount related to the spectrum envelope, and the other configurations and operations are the same as the components in FIG.

In the envelope enlarging unit 110, a codebook, which is a group of linear prediction coefficients having a wide band characteristic, is created in advance, and the linear prediction coefficients obtained from the analysis unit 102 are used as an input to use the vector quantization method. Then, a linear prediction coefficient having a wide band characteristic is selected from the above codebook, and this is used as the output of the envelope enlarging unit 110 and the input of the second speech synthesizing unit 109. Wide-band speech is synthesized from the generated spectral envelope feature values.

As described above, according to the voice band expanding apparatus of the present embodiment, the buffer 1 for storing the input signal for a certain period of time.
01, and an analysis unit 102 that extracts a feature amount related to the spectrum envelope from the signal sequence stored in the buffer.
And an envelope enlarging unit 110 that estimates a feature amount related to a spectral envelope having a wide band characteristic from a feature amount output from the analysis unit, and a feature amount from the analysis unit 102, with respect to a signal of the buffer 101. A pulse generator 107 that estimates a sound source pulse train and restores a missing band by using the estimated pitch period, and an enhancement filter 108 that weights the pulse train of the sound source with a frequency are provided. Then, it is possible to create a high-quality, wide-band synthetic sound with a widened band from the band-extended spectrum envelope and the band-expanded sound source.

In this embodiment, the first voice synthesizer 10
In the fifth and second speech synthesis units 109, the linear prediction coefficient is used as the feature amount to obtain the synthesized voice, but the PARCOR coefficient may be used instead of the linear prediction coefficient to perform the speech synthesis.

(Fourth Embodiment) Next, a fourth embodiment of the present invention will be described.

The overall construction of the fourth embodiment of the present invention is shown in FIG. The difference from the third embodiment is the configuration of the envelope enlarging unit, and the difference from the first embodiment is that the envelope enlarging unit 110 is provided as shown in FIG. 2 and the analyzing unit 102.
However, the point is that a linear prediction coefficient is calculated as the feature amount related to the spectrum envelope, and the other configurations and operations are the same as the components in FIG.

A detailed description of the envelope enlarging section 110 will be given with reference to FIG. FIG. 3 is an internal block diagram of the envelope enlarging unit 110. In FIG. 3, 301 is a narrowband codebook and 3 is a narrowband codebook.
Reference numeral 02 is a linear mapping function unit, and reference numeral 303 is a weighting addition unit. The input spectrum obtained by the analysis unit 102 is calculated with each code of the narrow band codebook 301 and the distance d _i calculated by (Equation 6).

[0058]

(Equation 6)

In (Equation 6), x _j is the j-th order input spectrum envelope information, and V _ij is the j-th order spectrum envelope information in the i-th code in the codebook 301.
Further, the input spectrum is converted by the linear mapping function unit 302 into a wide band spectrum by a plurality of linear mapping functions.
The output from the linear mapping function unit 302 is the weighting addition unit 30.
Weighted additions are made in 3 and output as a converted spectrum. The weight at that time is

[0060]

(Equation 7)

It is calculated by In (Equation 7), w _i is i
Is the weight for the output of the th linear mapping function. If each linear mapping function of the linear mapping function unit 302 is A _k ,

[0062]

(Equation 8)

The converted spectrum is calculated by In (Equation 8), y _j is the j-th order converted spectrum, and A j
_ij is the j-th function value of the i-th linear mapping function.

By using y _j thus obtained as the output of the envelope enlarging unit 110 and inputting it to the second speech synthesizing unit 109, from the synthesized sound source and the characteristic amount of the enlarged spectral envelope, Synthesize wideband speech.

As described above, according to the voice band expanding apparatus of the present embodiment, the buffer 1 for storing the input signal for a certain period of time.
01, and an analysis unit 102 that extracts a feature amount related to the spectrum envelope from the signal sequence stored in the buffer.
And an envelope enlarging unit 110 that estimates a feature amount related to a spectral envelope having a wide band characteristic from a feature amount output from the analysis unit, and a feature amount from the analysis unit 102, with respect to a signal of the buffer 101. A pulse generator 107 that estimates a sound source pulse train and restores a missing band by using the estimated pitch period, and an enhancement filter 108 that weights a frequency of the pulse train of the sound source are provided. With the configuration, it is possible to provide a voice band expanding device that can create a high-quality, wide-band synthetic sound with a band expanded from a spectrum envelope whose band is expanded with high precision and a sound source whose band has been expanded.

In this embodiment, the first voice synthesizer 10
In the fifth and second speech synthesis units 109, the linear prediction coefficient is used as the feature amount to obtain the synthesized voice, but the PARCOR coefficient may be used instead of the linear prediction coefficient to perform the speech synthesis.

[0067]

As described above, according to the present invention, a buffer for storing an input signal for a certain period of time is provided, and a feature quantity related to a spectrum envelope component is extracted from an audio signal in the buffer by an analysis unit, and the analysis is performed. From the output from the unit and the signal in the buffer, the sound source is represented by a plurality of pulse trains, the pulse train is weighted by distortion such as emphasis by the pitch period estimated by the pitch period estimation unit, and the pulse train that is the sound source is generated. By further performing a filtering process of weighting with respect to the frequency at the time of synthesis and synthesizing the voice from the feature amount of the spectrum envelope and the sound source pulse train, a relatively simple configuration,
It is possible to realize band expansion of a signal in a band not included in the narrow band signal without significantly impairing the characteristics of the original sound source.
In addition, a voice band expanding device and a voice band expanding device that generate a clear and wide band synthetic speech by adding a process of using the envelope expanding unit to widen the spectrum envelope information obtained by PARCOR analysis with high accuracy. A method can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing the overall configuration of a voice band expansion device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing the overall configuration of a voice band expansion device according to a third embodiment of the present invention.

FIG. 3 is a conceptual diagram of an envelope enlarging unit according to a fourth embodiment of the present invention.

FIG. 4 is a conceptual diagram of pulse enhancement in a pulse generator.

FIG. 5 is a block diagram showing the overall configuration of a voice band expansion device in a conventional example.

[Explanation of symbols]

 101 Buffer 102 Analysis Unit 103 Impulse Response Calculation Unit 104 Auditory Weighting Filter Unit 105 First Speech Synthesis Unit 106 Pitch Period Estimation Unit 107 Pulse Generation Unit 108 Enhancement Filter Unit 109 Second Speech Synthesis Unit 110 Envelope Enlargement Unit 201 LPC Analysis Part 202 Vector quantizer 203 Decoding part 204 Narrowband codebook 205 Wideband codebook 206 Low band restoration part 207 First high band restoration part 208 Second high band restoration part 209 Addition part 210 Upsampling part 301 Narrow band Codebook 302 Linear mapping function part 303 Weighting addition part

Claims (8)

[Claims] 1. A buffer for storing a fixed amount of an input signal,
An analysis unit that calculates at least a linear prediction coefficient of a linear prediction coefficient and a PARCOR coefficient for the signal sequence stored in the buffer, and an impulse response calculation that calculates an impulse response from the linear prediction coefficient obtained by the analysis unit. Section, an auditory weighting filter section for weighting a difference signal sequence between the output of the buffer and the output of the first speech synthesizer using the linear prediction coefficient as a parameter, and the auditory weighting filter section for simulating human auditory characteristics. A pitch period estimation unit that estimates a pitch period from the output signal of the weighting filter unit, and an output value from the auditory weighting filter and an output of the impulse response calculation unit with reference to the output value from the pitch period estimation unit. And a pulse train output from the pulse generator, and a linear prediction coefficient Of the first speech synthesizer for synthesizing a speech signal by inputting any one of the ARCOR coefficients, a frequency weighting filter for performing frequency weighting on the output of the pulse generator, and a frequency weighting filter. A voice band expanding device comprising a second voice synthesizing unit for synthesizing a voice by inputting an output value and one of the linear prediction coefficient and the PARCOR coefficient. 2. A buffer for storing a fixed amount of input signals,
An analysis unit that calculates at least a linear prediction coefficient of a linear prediction coefficient and a PARCOR coefficient for the signal sequence stored in the buffer, and an impulse response calculation that calculates an impulse response from the linear prediction coefficient obtained by the analysis unit. And a linear prediction coefficient or PARCOR whose band is expanded based on the linear prediction coefficient or PARCOR coefficient.
An envelope enlarging unit that estimates a coefficient, and an auditory weighting filter unit that weights a difference signal sequence between the buffer and the first speech synthesizer using the linear prediction coefficient as a parameter to simulate human auditory characteristics. With reference to the output value from the pitch period estimation unit that estimates the pitch period from the output signal of the auditory weighting filter unit, and the pitch period estimation unit,
A pulse generator that receives the output signal from the auditory weighting filter unit and the output from the impulse response calculator and generates a pulse train, a pulse train output from the pulse generator, the linear prediction coefficient and the PARCOR coefficient. And a frequency weighting filter for performing frequency weighting on the output of the pulse generating section, and an output value of the frequency weighting filter. And a second voice synthesizing unit for synthesizing voice by inputting an output value from the envelope enlarging unit. 3. A pulse generator has a frequency weighting filter for performing frequency weighting on an output signal from the auditory weighting filter, and has a function of generating a pulse train for restoring a missing band. The voice band expansion device according to claim 1 or 2, wherein. 4. The pulse generating section has a function of emphasizing the pulse by increasing the amplitude of the pulse at the pitch cycle interval using the detected value of the pitch cycle estimating section as a characteristic amount. Alternatively, the voice band expansion device described in 2. 5. The pulse generation section has a function of outputting a pulse set to the threshold value by using a certain threshold value with the detected value of the pitch period estimation section as a feature amount. The voice band expansion device described in 2. 6. The envelope enlarging unit has a function of inputting the feature amount obtained by the analyzing unit and widening the feature amount for the spectrum envelope having the feature of the wide band spectrum envelope using a linear mapping function. The voice band expansion device according to claim 2, wherein 7. A predetermined amount of input signal is stored in a buffer,
A feature amount related to the spectral envelope is extracted from the signal sequence stored in the buffer, and the source pulse sequence estimated from the extracted feature amount and the signal sequence stored in the buffer is a frequency of the signal. A voice band expanding method characterized by generating a wide-band source pulse train from a priori information about characteristics and estimated pitch information, and synthesizing a voice by using a feature amount related to the spectral envelope and the wide-band source pulse train. . 8. An input signal is stored in a buffer in a fixed amount,
A feature amount related to the spectral envelope is extracted from the signal sequence stored in the buffer, the envelope information missing in the envelope information represented by the feature amount is interpolated, and the feature amount and the signal stored in the buffer are interpolated. For a sound source pulse train estimated from the sequence, a wide-band source pulse train is generated from the a priori information and the estimated pitch information about the frequency characteristic of the signal, and the feature amount related to the spectrum envelope and the wide-band source pulse train. A method for expanding a voice band, which is characterized by synthesizing a voice using.