JPH0876799A - Wideband audio signal restoration method - Google Patents

Abstract

(57) [Summary] [Purpose] Restoring high-quality wide-band speech without nasal sounds with a small amount of calculation. [Structure] Narrowband speech (telephone speech) is subjected to LPC analysis, vector quantization is performed using a narrowband codebook 103, the quantization result is decoded by a wideband codebook 105 to obtain a wideband spectrum envelope, and this is LPC synthesized. Part 106
The LPC synthesis is performed in order to restore wideband speech. This is shown in JP-A-6-118995. An inverse filter 111 is created based on the LPC analysis result, a residual signal is created through narrow band speech, zero-filling is performed at every other point of the residual signal, and the zero-filled residual signal is used as the sound source signal of the LPC synthesis unit 106. To do.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for generating a wideband voice signal from a narrowband voice signal, and more specifically, a narrowband voice signal currently output by telephone voice or AM radio. , A method capable of improving the quality of a wideband audio signal such as that output from an audio set or FM radio.

[0002]

2. Description of the Related Art A case of telephone voice will be described as an example of a narrow band voice signal. The signal band that can be transmitted by the existing telephone system is about 300 Hz to 3.4 kHz. However, in recent years, there is an increasing number of voices who are not satisfied with the sound quality and desire high quality. As a method to solve this demand, it is possible to discard the existing telephone system and reconstruct a telephone system capable of transmitting a wideband signal, but it is not only a heavy burden on the economy but also the reconstruction. Even if it does, it will take a considerable amount of time. From such a point, as seen in Japanese Patent Laid-Open No. 6-118995, a signal which exists in a wide band signal but does not exist in a narrow band signal is restored from the narrow band signal.
It has been proposed to generate wideband signals. This method is advantageous because wideband voice can be obtained regardless of the configuration of the telephone network.

The method disclosed in the above publication will be briefly described with reference to FIG. Input narrow band audio signal is LPC
LPC analysis is performed in the analysis unit 101, and the analysis result is fuzzy vector quantized in the quantization unit 102 using the codebook 103 of the narrowband speech signal. For fuzzy vector quantization, see H. Tseng,
M. Sabin, E .; Lee, “Fuzzy Vect
or Quantization Applied t
o Hidden Markov Modelin
g, "ICASSP'87 15.5, Apr. 198.
7. Are described in. Quantization unit 3 to reduce calculation amount
The processing of 02 may be ordinary vector quantization. The fuzzy vector quantized code is decoded by the decoding unit 104 using the codebook 105 of the wideband speech signal.

The decoded output X'is a speech spectrum envelope, which is used by the speech synthesis unit 106 in the LPC.
A wideband speech signal is obtained by synthesizing. That is, the synthesis filter is driven by the excitation signal corresponding to the pitch obtained from the analysis result of the LPC analysis unit 101, the filter coefficient of the synthesis filter is controlled according to the decoded output X ′, and the LPC analysis unit 101 is controlled.
A synthesized voice signal having a level corresponding to the power obtained from the analysis result of is obtained. The synthesized voice signal may be output as a restored wideband voice signal.

The wideband speech signal obtained by the above processing is
Although a signal that does not exist in the input narrowband voice signal is included in the band to form a wideband signal, a signal different from the input narrowband voice signal is also included in the band of the input narrowband voice signal. If the audio signal is used as it is, the side effect of distorting the signal existing in the input narrow band audio signal occurs. Therefore, the following processing is performed and the signal existing in the original input narrow band speech signal is used as it is. That is, the wideband speech signal synthesized by the LPC synthesis section 106 is supplied to the bandpass filter 107, and out-of-band components of the input narrowband speech signal, that is, frequency components of 300 Hz or less and frequency components of 3.4 kHz or more are extracted. Be done. On the other hand, the input narrow band audio signal is up-sampled in the 8 kHz band by the up-sampling unit 108. The output of the up-sampling unit 108 and the output of the band-pass filter 107 are added by the adding unit 109,
A reconstructed wideband speech signal is obtained. In the up-sampling, for example, zero samples are inserted between the sample points and passed through an all-pass filter, and the output is sampled at a double speed to double the frequency band.

A codebook 10 for wideband speech signals
5 is created using a large number of learning wideband speech signals by a method similar to that of a codebook used for vector quantization of speech, for example. In addition, the codebook 103 for narrowband speech signals is the codebook 1 for wideband speech signals.
It was made in association with No. 05, and its making method is described in the above-mentioned publication.

[0007]

In the method disclosed in the above publication, in order to generate the out-of-band component of the input narrowband speech signal, the LPC synthesis section 106 obtains the linear predictive speech synthesis filter in the LPC analysis section 101. It is driven by an impulse train having a given pitch period, that is, the impulse train is used as a sound source signal, and the synthesized waveform is transformed by the wideband spectrum envelope decoded by the decoding unit 104.

[0008] Synthetic speech by an impulse train has a nasal timbre because energy is concentrated on the waveform, and it cannot be said to be high-quality synthetic speech. Therefore, the deformation that disperses the energy is necessary. This modification is effective in removing the abnormal noise in the composite waveform, but on the other hand, it increases the amount of calculation. An object of the present invention is to provide a wideband speech signal restoration method capable of removing abnormal noise and obtaining a high-quality wideband speech signal with a small amount of calculation.

[0009]

According to the invention of claim 1, a residual signal is created from an input narrow band speech signal, and the residual signal is padded with zeros to have the same band as the wide band speech signal.
A wideband speech signal is restored using the zero-padded residual signal as a sound source signal. According to the second aspect of the present invention, a residual signal is created in advance from the wide band speech signal, and the wide band speech signal is restored using this residual signal as the sound source signal.

[0010]

In contrast to the impulse, the residual waveform does not have nasal sound quality because the energy is dispersed appropriately. Furthermore, the residual waveform can be determined by using the inverse characteristic of the linear predictive speech synthesis filter, so that its spectrum is white. Therefore, if a high-frequency spectrum envelope is given in some form, it is possible to generate a speech signal having the spectrum feature as it is.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the invention of claim 1 will be described with reference to FIG. In FIG. 1, parts corresponding to those in FIG. 3 are designated by the same reference numerals. In the present invention, the LPC coefficient obtained by the LPC analysis unit 101 is used to form an inverse filter 111, and a narrowband speech signal is input to the inverse filter 111 to obtain its residual signal. The residual signal is zero-padded by the padding processing unit 112 to generate a sound source signal for wideband speech synthesis.
For example, in order to create a sound source signal in the 8 kHz band from the residual signal in the 4 kHz band, zero is padded for each point. This zero-filled residual signal is used as the source signal L
The signal is supplied to the PC synthesizer 106 to obtain a wideband voice signal by the LPC synthesis method. Although this wideband audio signal may be used as a restored wideband audio signal as it is, it is better to use only the low-frequency component and the high-frequency component as in the prior art and to use the input narrow-band audio as the component between them. .

In the above embodiment, the residual signal is obtained from the input narrow band voice signal each time, and zero is padded to the sound source signal. However, the residual signal is obtained in advance from the wide band voice signal. You can also FIG. 2A shows how to obtain the residual signal. The analysis unit 201 performs LPC analysis on the wideband speech signal, and an LPC coefficient obtained by the analysis is used to configure an inverse filter 202.
A wideband speech signal is input to 2 to obtain a residual signal. The residual signal is input to the voiced / unvoiced determination unit 203 to distinguish between voiced sound and unvoiced sound of the input band speech signal. As a residual signal and an unvoiced residual signal, the memory 204,
Each is stored in 205. The spectrum of the residual signal is
Since it has a flat characteristic, it can be commonly used for all phonemes without considering which phoneme the residual signal is obtained from, except for the discrimination between voiced sound and unvoiced sound. The residual signal to be stored is at least 1 for voiced sounds.
It is enough for the basic frequency and one frame for unvoiced sound.

The invention of claim 2 obtains a wideband speech signal by using the residual signal prepared in advance as a sound source signal. An embodiment thereof is the same as the portion corresponding to FIGS. 1 and 2A in FIG. 2B. Shown with a reference numeral. In this embodiment, the voiced / unvoiced determination unit 206 determines voiced sound and unvoiced sound by using the LPC analysis result of the input narrow band speech signal, and the sound source signal generation unit 207 stores the memory 204,
The residual signal is taken out from 205 and a sound source signal is generated.
In the case of voiced sound, the voiced residual signal extracted from the memory 204 is arranged for each basic period. If the voiced residual signal is short of the fundamental period, it is padded with zeros. If the voiced residual signal is longer than the fundamental period, the voiced residual signal is truncated before use. In the case of unvoiced sound, the unvoiced residual signal of one frame is used as it is. The sound source signal generated in this manner is supplied to the LPC synthesizer 106 to synthesize voice.

The spectrum envelope of wideband speech used in the LPC synthesizer 106 may be obtained by another method.

[0015]

As described above, according to the present invention, the residual signal, not the impulse train, is used as the sound source signal for wideband speech synthesis, so that the transformation of the synthesized waveform is unnecessary and the amount of calculation is greatly increased. Can be reduced. Moreover, a high-quality wideband audio signal is obtained without nasal sound quality.

[Brief description of drawings]

FIG. 1 is a flowchart showing a procedure of an embodiment of the invention of claim 1.

2A is a flow chart showing a procedure for obtaining a residual signal from a wideband speech signal, and B is a flow chart showing a procedure of an embodiment of the invention of claim 2. FIG.

FIG. 3 is a flowchart showing a procedure of a conventional wideband speech restoration method.

Claims (4)

[Claims] 1. A method of generating a wideband speech signal from an input narrowband speech signal by using a method of speech-synthesizing a speech signal with a spectrum envelope and a sound source signal, wherein the residual signal from the input narrowband speech signal is used. In order to restore the wideband speech signal by filling the residual signal with zeros by an amount corresponding to the difference between the frequency bands of the input narrowband speech signal and the wideband speech signal, and zeroing the residual signal. Wideband speech signal restoration method characterized by being used as a sound source signal of a. 2. A method for recovering a wideband speech signal from an input narrowband speech signal using a method of speech-synthesizing a speech signal into a spectrum envelope and a sound source signal, wherein a residual signal is obtained from the wideband speech signal. A wideband speech signal restoration method characterized in that the residual signal is stored and used as a sound source signal for restoring the wideband speech signal. 3. A residual signal of voiced sound and a residual signal of unvoiced sound are distinguished and stored as the residual signal, and it is determined whether the input narrowband speech signal is voiced sound or unvoiced sound. 3. The wideband speech signal restoration method according to claim 2, wherein the voiced sound residual signal is used in the case of a voice sound, and the unvoiced residual signal is used in the case of an unvoiced sound as the sound source signals. 4. An out-of-band component of the input narrowband speech signal is extracted from a signal obtained by speech synthesis using the sound source signal,
4. A wideband audio signal restored by adding this and the input narrowband audio signal to obtain a wideband audio signal.
A wideband audio signal restoration method according to any one of 1.