JP2000122679A - Audio range expanding method and device, and speech synthesizing method and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the number of wide band formants to lay stress on rough structure of spectra, to enhance the quality of a provided wide audio range voice thereby, and to save a memory capacity and an operational amount required for code book search. SOLUTION: In this audio range expanding device, a linear prediction coefficient α is converted into an auto-correlation (r) in a parameter converting part 7, a wide range auto-correlation rw is generated from the auto-correlation (r) in a vector quantizing part 8 and a vector dequantizing part 9, the auto- correlation rw is converted reversely into a wide range linear prediction coefficient αw in a parameter reversely converting part 10, and a wide range voice is synthesized from the prediction coefficient αw in an LPC synthesizing part 11.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for transmitting an audio signal having a narrow frequency band or a parameter constituting the audio signal transmitted or stored in a medium, for example, in a transmission, a transmission path, or a medium. A voice band extending method and apparatus, and a voice synthesizing method, which are used in a case where a broadband voice signal is estimated on a receiving side or a reproducing side, and particularly suitable for a mobile phone terminal having a band extending function, etc. And an apparatus.

[0002]

2. Description of the Related Art Conventionally, as a band extension technology, a voice codec system of a mobile phone or a mobile phone based on a PDC (Personal Digital Cellular) system in Japan, which uses a frequency band of 300 Hz to 3400 Hz, is used.
SELP (Vector Sum Excited Linear Prediction) and PSI-CELP (Pitch Synchronous Innovation)
-Estimating out-of-band signal components on the receiving side from audio signals of the Code Excited Linear Prediction) method,
There is a technique for widening the band to about Hz to 6000 Hz.

In this technique, a wideband parameter is estimated from a transmitted parameter, and a wideband speech is synthesized using the obtained wideband parameter. Further, since the signal component of 300 Hz to 3400 Hz included in the original audio signal has less distortion than the band component of 300 Hz to 3400 Hz included in the synthesized audio signal, for the purpose of further improving the sound quality, For example, the band component of 300 Hz to 3400 Hz is removed from the synthesized sound signal by a filter or the like, and the band component of 300 Hz to 3400 Hz of the original audio signal is added to the synthesized sound signal after the band removal.

The transmitted narrow band parameters include a linear prediction coefficient (α), a reflection coefficient (k), and a line spectrum pair (LSP). These represent the spectral envelope of the voice, and the order (number) of the coefficients corresponds to the peak of the spectrum. In the PDC system, the coefficient is transmitted up to the tenth order because of the nature that human voices up to about 3400 Hz have about five formants.

There are various methods for predicting a wideband parameter representing a wideband formant.
One is a method using vector quantization. This method
A plurality of order dimension vectors of the wideband parameters are prepared by learning in advance, and when the narrowband parameters are input, an appropriate wideband vector is selected from the input to obtain the wideband parameters.

[0006]

When estimating and synthesizing a wideband speech from a transmitted narrowband parameter, the number of formants naturally exceeds the narrowband number of five.

However, as the order of the parameter representing the wideband formant increases, the dimension of the vector at the time of the above-described vector quantization increases, thereby first increasing the memory capacity, and processing required for the codebook search. The amount also increases.

In addition, the fact that the number of formants increases means that comparison is made up to the fine structure of the spectral envelope, which is far from the purpose of estimating the rough broadband spectral envelope which was originally important.
There are few benefits.

Accordingly, the present invention has been made in view of such a situation, and it is possible to reduce the number of broadband formants, attach importance to the rough structure of the spectrum, and improve the quality of the obtained wideband speech. It is another object of the present invention to provide a voice band expansion method and apparatus, and a voice synthesis method and apparatus, which can reduce the memory capacity and the amount of calculation required for codebook search.

[0010]

SUMMARY OF THE INVENTION A voice band extending method according to the present invention includes a parameter extracting step of obtaining a parameter capable of expressing a narrow band formant from an input narrow band voice signal, and a method of obtaining a parameter which is smaller than the obtained number of narrow band formants. The object is achieved by estimating parameters that can represent not a large number of wideband formants and synthesizing wideband speech from the parameters that represent the obtained wideband formants.

The voice band extending apparatus according to the present invention comprises a parameter extracting means for obtaining a parameter capable of expressing a narrow band formant from an input narrow band voice signal, and a wide band formant having a number not larger than the number of obtained narrow band formants. The above-mentioned problem is solved by having parameter prediction means for predicting a parameter capable of expressing a wideband formant and synthesis means for synthesizing a wideband speech from the obtained parameter representing a wideband formant.

According to the speech synthesis method of the present invention, a number of broadband formants that are not more than the number of narrowband formants are selected from parameters that can represent narrowband formant information among narrowband parameters representing input narrowband speech. A first parameter prediction step of predicting a parameter that can be expressed, a parameter extraction step of obtaining a parameter that can express narrowband formant information from the input narrowband speech, and a number not more than the obtained number of narrowband formants The above-described problem is solved by having a second parameter prediction step of predicting a parameter capable of expressing the wideband formant of the above and a synthesis step of synthesizing a wideband speech from the obtained parameter expressing the wideband formant.

The speech synthesizing apparatus according to the present invention, from among the narrow-band parameters representing the narrow-band speech to be input, parameters that can represent narrow-band formant information, extracts a number of wide-band formants not greater than the number of narrow-band formants. First parameter prediction means for predicting a parameter that can be expressed, parameter extraction means for obtaining a parameter capable of expressing narrowband formant information from the input narrowband speech, and a number not more than the obtained number of narrowband formants The above-mentioned problem is solved by having second parameter prediction means for predicting a parameter capable of expressing the wideband formant and synthesis means for synthesizing a wideband speech from the obtained parameter representing the wideband formant.

That is, according to the present invention, the number of formants to be originally estimated has to be larger than the number of formants in the narrow band. Only rough structure is estimated.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described with reference to the drawings.

As an embodiment of the voice band extension method and apparatus and voice synthesis method and apparatus of the present invention, an example will be described in which a VSELP system and a PSI-CELP system, which are PDC codec systems, are applied.

In this embodiment, a wideband parameter is estimated from a narrowband parameter, and a wideband LPC synthesis is performed.
After that, the low-frequency side of the frequency band of the original audio is replaced with the original audio signal. That is, in the present embodiment, high-pass filtering is performed on the synthesized voice signal to leave only the high frequency band, high frequency components among the high frequency components are suppressed, the gain is further adjusted, and added to the original voice signal. Is performed.

Here, the estimation of the wideband parameter requires two steps: a broadening of the linear prediction coefficient α, which is a parameter representing the spectrum envelope, that is, formant information, and a widening of the excitation source. Further, in order to increase the bandwidth of the linear prediction coefficient α, it is necessary to create a codebook based on the autocorrelation r which is a parameter that can be mutually converted with the linear prediction coefficient α. Quantization by this codebook,
The autocorrelation r is broadened by inverse quantization.

Hereinafter, referring to both FIG. 1 and FIG. 2, the extension of the linear prediction coefficient α, the extension of the excitation source, and the broadband LP
We will explain along the flow of processing of C synthesis and low-frequency replacement,
Then, how to create a codebook is described. Figure 1 shows the PSI
FIG. 2 shows a block diagram of the present embodiment when the CELP scheme is applied, and FIG. 2 shows a block diagram of the present embodiment when the VSELP scheme is applied. In FIGS. 1 and 2, parts performing the same processing are denoted by the same reference numerals.

First, the expansion of the linear prediction coefficient α will be described.

In the present embodiment, attention is paid to the fact that the linear prediction coefficient α is a filter coefficient representing a spectrum envelope, and the autocorrelation parameter representing another spectrum envelope which makes it easy to estimate the high-frequency side by the parameter converter 7. once converted to r, then it is broadband, further followed to inverse transformation by the parameter inverse conversion unit 10 from the wide-band autocorrelation r _W in wideband linear prediction coefficient alpha _W.

At this time, vector quantization is used to extend the autocorrelation r (to increase the bandwidth). That is, the narrow band autocorrelation r _N and the vector quantization by the vector quantization unit 8, by vector inverse quantization of the index at vector inverse quantization section 9, the wide band autocorrelation r _W corresponding from the index Ask for it.

Since a certain relationship is established between the narrowband autocorrelation and the wideband autocorrelation as described later, only a codebook based on the wideband autocorrelation needs to be prepared, and the narrowband autocorrelation can be vector-quantized by this. Wideband autocorrelation is obtained by vector inverse quantization.

Here, assuming that the narrow-band signal is obtained by band-limiting the wide-band signal, the wide-band autocorrelation and the narrow-band autocorrelation have the following relationship.

[0025]

(Equation 1)

Here, φ is the autocorrelation, _xn is the narrowband signal, _xw is the wideband signal, and h is the impulse response of the band-limiting filter.

Further, the following equation is obtained from the relationship between the autocorrelation and the power spectrum.

[0028]

(Equation 2)

If another band-limiting filter having a frequency characteristic equal to the power characteristic of this band-limiting filter is considered and this is H ′, the following equation is obtained.

[0030]

(Equation 3)

The pass band and the stop band of this new filter are the same as those of the original band limiting filter, and the attenuation characteristic is squared. Therefore, this new filter can also be said to be a band limiting filter.

Taking this into account, the narrow-band autocorrelation is simplified to the convolution of the wideband autocorrelation with the impulse response of the band-limiting filter, that is, the band-limited version of the wideband autocorrelation. That is, the following equation is obtained.

[0033]

(Equation 4)

As described above, in the vector quantization of the narrowband autocorrelation, if only a wideband codebook is prepared, the narrowband vector required at the time of quantization can be created by calculation. It is not necessary to prepare a code book in advance.

Further, since each broadband autocorrelation r _W code vector has a curve that monotonically decreases or gradually increases and decreases, there is no significant change even when the bandpass filter H ′ passes the low band, and the narrowband autocorrelation r _N quantum vector does not change. The conversion can also be performed directly with the wideband autocorrelation r _W codebook. However, since the sampling frequency is 1/2, r
_It is necessary to compare the _W code vector with the first order and r _N.

Here, the autocorrelation parameter is PD
In the case of C, it is possible to obtain up to the 10th order in a narrow band, but there is a property that the lower order one expresses a rough spectral envelope, and the higher order one expresses a finer structure. Therefore,
Autocorrelation is naturally required up to about the 20th order in wideband speech with increased sampling frequency. In the present embodiment, however, in order to emphasize rough spectral envelope and to reduce the amount of calculation and memory capacity, for example, the 6th order is used. I will only ask to the extent. Therefore, the dimension of the wideband codebook is six in this case.

The linear prediction coefficient α can be extended into voiced sound (V) and unvoiced sound (UV) more accurately by dividing it into voiced (V) and unvoiced (UV) sounds. That is, in the present embodiment, the V / UV discriminating unit 2 discriminates the decoded audio signal, and performs processing according to the discrimination result. Along with this, two codebooks for the voiced sound (V) and unvoiced sound (UV) are also used in the codebooks in the vector quantization unit 8 and the vector inverse quantization unit 9.

Next, expansion of the excitation source will be described.

In the PSI-CELP method used in FIG. 1, the excitation source in a narrow band is upsampled by inserting a zero value in a zero padding section 20 to generate aliasing distortion. . Although this method is very simple, it can be said that the quality is sufficient as an excitation source because the difference between the power and the harmonic structure of the original voice is preserved.

However, the VSELP used in FIG.
In the method, the vowels of the original voice are cloudy. If the method of inserting a zero value into the excitation source is applied as it is, unpleasant noise remains in the high frequency range. To improve this, FIG.
In the present embodiment, the following processing is performed.

The excitation source of the VSELP method shown in FIG. 2 includes parameters β (long-term prediction coefficients) and bL used for the codec.
[i] (long-term filter state), γl (gain), cl [i] (excitation code vector), β * bL [i] + γl * cl
[i], of which the former is the pitch component,
Since the latter represents a noise component, in the example of FIG.
* BL [i] (first excitation source E1) and γl * cl [i] (second excitation source E2).
These energies are compared for each sub-frame. If the energy of the former (first excitation source E1) is large, only the pitch component is emphasized and the excitation source is set to a pulse train. It is detected whether the sample value of the source E1 is equal to or more than a predetermined value, that is, whether there is a pitch component. The sample value of the first excitation source E1 is used in a portion having a pitch component, and 0 in a portion having no pitch component. To suppress.
Further, the first excitation source E
If the energy of the first excitation source E1 is not greater than the energy of the second excitation source E2,
The one to which the excitation source E2 is added is used. The narrow band excitation source thus created is padded with zero values by the zero padding unit 21 in the same manner as in the PSI-CELP method, thereby generating a wide band excitation source. If this processing is written in C-like, it becomes as follows.

[0042]

(Equation 5)

Next, as wideband LPC synthesis, the LPC synthesis section 11 performs LPC synthesis using the wideband linear prediction coefficient α obtained above and a wideband excitation source.

Next, low-frequency replacement will be described.

The speech that has been subjected to wideband LPC synthesis by the LPC synthesis unit 11 is accompanied by estimation errors, including a reduction in the number of formants, and is of poor quality if left unaltered. It is to be replaced with the original audio signal SND _N outputs. For this reason, in the present embodiment, 4 kHz or more of the synthesized sound signal from the LPC synthesis unit 11 is extracted by the narrowband frequency range elimination unit 12, while the codec output is fs = 16 kHz by the up-sampling unit 5. And these are added by the adder 17.

At this time, in the present embodiment, the high-frequency gain can be adjusted as desired. That is, since the sound quality greatly varies among users, it is important to make the value of the high-frequency gain variable. Therefore, in the present embodiment, the value of the high-frequency side gain is set in advance by an input from the user, and the multiplication unit 16 is referred to by referring to this gain value.
By multiplying the gain value, the high-frequency gain is adjusted.

In the present embodiment, the high-band suppression filter unit 13 performs a filtering process for slightly suppressing components of about 6 kHz or more on the high-frequency side signal before addition in the addition unit 17. The sound is easy to hear. This filter coefficient is selectable, and by performing a filtering process using a filter coefficient selected in advance, it becomes possible to select a high-frequency band according to preference. The selection of this filter is also set by the user's input.

However, the high band suppression filter processing by the high band suppression filter unit 13 may be performed after the addition by the addition unit 17 because it does not affect the power characteristics on the low band side. Alternatively, it is also possible to apply a filtering process that also affects the low-frequency side after the addition by the adding unit 17.

As described above, a wideband voice can be obtained.

Next, a method of creating a code book will be described.

In this embodiment, a code book is created before the above-described band expansion processing. FIG. 3 shows a block diagram for generating codebook training data, and FIG. 4 shows a block diagram for generating a codebook.

The code book is created using a well-known generalized Lloyd algorithm (GLA).

In FIG. 3, in the present embodiment, first, the framing section 31 converts a wideband audio signal into a signal for a predetermined time.
For example, it is divided into frames every 20 ms, and V / UV
The classification unit 32 classifies the voiced sound (V) and the unvoiced sound (UV), and for each frame of the voiced sound (V) and the unvoiced sound (UV), the autocorrelation calculation units 33 and 34 respectively set a fixed order,
For example, an autocorrelation up to the sixth order is obtained. The autocorrelation r for each frame of the voiced sound (V) and the unvoiced sound (UV) is used as training data.

Next, referring to FIG. 4, in the present embodiment,
A wideband parameter extracting unit 41 extracts a wideband parameter from the autocorrelation r of each of the voiced sound (V) and unvoiced sound (UV) for each frame, and a codebook learning unit 42 creates a six-dimensional codebook. I do.

As described above, the voiced sound and the unvoiced sound are distinguished from each other, and the autocorrelation of the voiced sound and the autocorrelation of the unvoiced sound are separately collected. Reference is made to a book. At this time, a voiced sound or an unvoiced sound is determined, and a corresponding code book is used.

It should be noted that the codebook may be created without distinguishing voiced and unvoiced sounds.

As described above, according to the present embodiment, by reducing the number of broadband formants,
Emphasis was placed on the rough structure of the spectrum, and the quality of the resulting wideband speech was improved. Further, the memory capacity and the amount of calculation required for the codebook search can be saved.

The parameters that can represent the formant are not limited to the linear prediction coefficient α and the autocorrelation r described in the present embodiment, but may be, for example, a line spectrum pair (LS
Various types such as P) and a partial autocorrelation coefficient (PARCOR coefficient) can be used. In addition, the present invention is not limited to only predicting the high band from the low band, and is not limited to the PDC method. Furthermore, even if analog voice is transmitted, the present invention can be used as long as digital signal processing is performed thereafter. Therefore, the present invention is not limited to parameter transmission. Still further, the present invention can be applied in exactly the same manner to those that do not pass through a transmission path, such as an answering machine, a response message, and memo recording, which are functions of a portable terminal.

[0059]

As is apparent from the above description, in the present invention, in a band extension technique for predicting and synthesizing a wideband speech from a narrowband speech or a narrowband parameter, the formant of the synthesized wideband speech is reduced. The rough structure of the spectrum can be emphasized, the quality of the obtained wideband speech can be improved, and the memory capacity and the amount of calculation required for a codebook search can be reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment when a PSI-CELP scheme is applied to the present invention.

FIG. 2 is a block diagram of an embodiment when the VSELP method is applied to the present invention.

FIG. 3 is a block diagram for generating training data.

FIG. 4 is a block diagram for generating a code book.

[Explanation of symbols]

2 V / UV discriminating unit, 4 first order unit, 5 up-sampling unit, 7 parameter conversion unit, 8 vector quantization unit, 9 vector inverse quantization unit, 10 parameter inverse conversion unit, 11 LPC synthesis unit, 12 narrow band Frequency range remover, 13 high-frequency suppression filter, 16
Multiplying unit, 17 adding unit, 20, 21 Zero padding unit, 22 Frame energy comparing unit, 23 Pitch component detecting unit, 31 Framing unit, 32 V / U
V classification unit, 33,34 autocorrelation calculation unit, 41 wideband parameter extraction unit, 42 wideband codebook learning unit

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] December 9, 1998 (1998.12.
9)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claim 5

[Correction method] Change

[Correction contents]

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] Claim 7

[Correction method] Change

[Correction contents]

[Procedure amendment 3]

[Document name to be amended] Statement

[Correction target item name] 0012

[Correction method] Change

[Correction contents]

According to the speech synthesis method of the present invention, a number of broadband formants that are not more than the number of narrowband formants are selected from parameters that can represent narrowband formant information among narrowband parameters representing input narrowband speech. The above-described problem is solved by having a parameter prediction step of predicting a parameter that can be expressed and a synthesis step of synthesizing a wideband speech from the obtained parameter expressing the wideband formant.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0013

[Correction method] Change

[Correction contents]

The speech synthesizing apparatus according to the present invention, from among the narrow-band parameters representing the narrow-band speech to be input, parameters that can represent narrow-band formant information, extracts a number of wide-band formants not greater than the number of narrow-band formants. Parameter prediction means for predicting a parameter that can be represented;
Having synthesis means for synthesizing a wideband speech from the parameters expressing the obtained wideband formant,
The above-mentioned problem is solved.

[Procedure amendment 5]

[Document name to be amended] Statement

[Correction target item name] 0042

[Correction method] Change

[Correction contents]

[0042]

(Equation 5)

Claims (8)

[Claims] 1. A parameter extracting step for obtaining a parameter capable of expressing a narrow-band formant from an input narrow-band audio signal, and predicting a parameter capable of expressing a number of wide-band formants which is not more than the number of obtained narrow-band formants. And a synthesizing step of synthesizing a wideband speech from a parameter expressing the obtained wideband formant. 2. The voice band according to claim 1, further comprising a replacement step of removing a frequency range corresponding to the narrow-band voice signal from the synthesized wide-band voice signal and replacing it with the input narrow-band voice signal. Expansion method. 3. A parameter extracting means for obtaining a parameter capable of expressing a narrow-band formant from an input narrow-band audio signal, and predicting a parameter capable of expressing a number of wide-band formants which is not more than the obtained number of narrow-band formants. And a synthesizing unit for synthesizing a wideband speech from a parameter expressing the obtained wideband formant. 4. The voice band according to claim 3, further comprising a replacement unit for removing a frequency range corresponding to the narrow band voice signal from the synthesized wide band voice signal and replacing the frequency range with the input narrow band voice signal. Expansion device. 5. A parameter that can express a number of wideband formants that is not more than the number of narrowband formants from parameters that can express narrowband formant information among narrowband parameters that represent input narrowband speech. A first parameter prediction step of predicting, a parameter extraction step of obtaining a parameter capable of expressing narrowband formant information from an input narrowband speech, and expressing a number of wideband formants not more than the obtained number of narrowband formants Second to predict possible parameters
And a synthesizing step of synthesizing a wideband speech from the parameters representing the obtained wideband formants. 6. The speech synthesis according to claim 5, further comprising a replacement step of removing a frequency range corresponding to the narrowband speech signal from the synthesized wideband speech signal and replacing the frequency range with the input narrowband speech signal. Method. 7. A parameter that can express the number of wideband formants that is not more than the number of narrowband formants from parameters that can express narrowband formant information among narrowband parameters that represent input narrowband speech. First parameter predicting means for predicting, parameter extracting means for obtaining a parameter capable of expressing narrowband formant information from the input narrowband speech, and expressing the number of wideband formants not more than the number of obtained narrowband formants Second to predict possible parameters
And a synthesizing unit for synthesizing a wideband speech from the parameters representing the obtained wideband formants. 8. A speech synthesizer according to claim 7, further comprising replacement means for removing a frequency range corresponding to a narrowband speech signal from the synthesized wideband speech signal and replacing the frequency range with an input narrowband speech signal. apparatus.