JPH06208398A - Generation method for sound source waveform - Google Patents

Abstract

PURPOSE:To improve the quality of a reproduced sound of the heading and the ending of a word by changing an amplitude of a sound source code book in the same way as amplitude change in an input sound, when frame power of the input sound has a rising tendency or a falling tendency over the range from the frame to thee next frame. CONSTITUTION:A rising and falling detecting circuit 28 detects the rising of an input sound by discriminating a threshold value, and also detects the inclination of rising. A sound source waveform generation circuit 30 transists its output amplitude in a frame according to this inclination. A signal level outputted from a sound source code book 10 is constant in the same frame, and the gain B of a coefficient adder 14 is set in an optimum state by an error power minimizing section 26. Therefore, since the sound source waveform is weighted by that inclination, the appropriate sound source waveform is obtained by using this processing even if an input sound is raised or fallen with shorter lag than frame length.

Description

Detailed Description of the Invention

[0001]

The present invention relates to CELP (code excit
ed linear prediction: a method for generating a sound source waveform of a sound source codebook in a code excitation linear prediction) coding method.

[0002]

2. Description of the Related Art For example, a voice CODEC used in a digital car telephone is required to reduce a coding rate of a digital voice signal from 4 kbps to 8 kbps. FIG. 6 shows an example of a method of generating a sound source waveform used for such a purpose.

The method shown in this figure is a method of generating an excitation waveform with an excitation codebook in the CELP coding method. In the figure, reference numeral 10 denotes a sound source codebook, and the sound source codebook 10 stores a signal used as a sound source signal in a past frame. That is, a linear configuration including a coefficient multiplier 14 with a gain β (scalar), a coefficient multiplier 16 with a gain γ (scalar), and an adder 18 for performing vector addition, which is provided at the subsequent stage of the sound source codebook 10 and the noise codebook 12. The output of the adding means is used as a sound source signal in the current frame and is fed back to the sound source codebook 10.

The frame referred to here is a sound signal used as a sound source signal, which is framed with a length of about 20 msec, and has a sampling frequency of 8 kHz.
When z is used, the number of samples belonging to one frame is 8 kHz × 20 msec = 160 samples. It is necessary to set the frame length to a length at which the audio signal can be considered to be stationary, that is, to a length at which the property of the audio signal does not change significantly, and is usually set to about 20 to 30 msec. The sound source codebook 10 discards the oldest sound source signal and sequentially stores new sound source signals to be fed back (update of the sound source codebook 10). As a result, the sound source codebook 10 stores sound source signals of past frames, such as a signal used as a sound source signal in the previous frame, a signal used as a sound source signal in the previous frame, and so on. . At this time, one frame is divided into four subframes and handled. That is, the sound source codebook 10 is updated in units of this subframe, that is, in units of 160/4 = 40 samples, so that the newest subframe is located at the head.

The audio signals of, for example, 160 samples stored in the sound source codebook 10 are output to the linear adding means in a state in which several tens of samples (40 samples in the above example) are put together. A signal obtained by collecting dozens of samples is called a vector. In particular, the vector output from the sound source codebook 10 is called a sound source code vector.

The sound source code vector required for optimal sound source signal generation is A-B-S (analysis by synthesis).
sis: Synthetic analysis) method. A-B-S
The method is a method of actually combining and adding feedback using a closed loop in order to generate a signal with the smallest error power. The A-B-S method in FIG. 6 is mainly realized by determining the optimum lag in the excitation codebook 10 and determining the optimum index in the noise codebook 12.

First, the optimum lag in the excitation codebook 10 is determined by searching the lag in the excitation codebook 10 based on the error power of the output of the weighting synthesis filter 20 with respect to the output of the auditory weighting filter 22. The search for each vector in the sound source codebook 10 and the noise codebook 12 is performed independently.

As described above, the excitation code vector output from the excitation code book 10 is linearly added to the output of the noise code book 12 by the coefficient multipliers 14 and 16 and the adder 18, and then weighted as an excitation signal. It is input to the synthesis filter 20. When searching the sound source code vector, the input from the noise codebook 12 is set to zero. The weighting synthesis filter 20 is a filter for converting the sound source signal into a form that can be compared with the output of the auditory weighting filter 22, and applies a predetermined weight to the output of the adder 18.
On the other hand, the input voice (voice to be encoded) is weighted by the auditory weighting filter 22. Subtractor 24
Calculates the error power of the output of the weighting synthesis filter 20 with respect to the output of the auditory weighting filter 22 by vector subtraction. The error power minimization unit 26 provided in the subsequent stage of the subtractor 24 minimizes the obtained error power.
A search in the sound source codebook 10 is executed. The error power minimization unit 26 searches a predetermined search range so that the output of the weighting synthesis filter 20 becomes a vector most similar to the output of the auditory weighting filter 22. The lag with the minimum error power is called the optimum lag.

[0009] Here, the lag is the tone generator codebook 10
The age of the sample as seen from the newest sample. As described above, the sound source codebook 10 stores signals that have been used as sound source signals in the past.
If the search range is, for example, −20 to −146 samples (from 20 samples before to 146 samples before the newest sample in the sound source codebook 10), the error power minimization unit 26 selects any of the samples belonging to this search range. Forty samples (one vector) starting from the sample are sequentially taken out, and a search for the minimum error power (search for the optimum lag) is performed over the entire search range. The position of the sample that is the starting point is called the lag, and the lag of the vector that minimizes the error power is called the optimum lag. The lag corresponds to the cycle of vocal cord vibration, and the optimum lag is longer when the input voice is a male voice and shorter when the input voice is a female voice. The search range is set narrow when the coding speed is lowered (the bit rate is lowered).

The optimum lag obtained by performing the search in this way can also be used at the time of decoding. That is, the excitation codebook 10 is encoded by the same method on the encoding side and the decoding side.
Is updated, it is possible to give the information related to the excitation codebook 10 only by transmitting the information of the optimum lag from the encoding side to the decoding side. Further, in the above-described search range setting example, the lag can be represented by 7 bits.

In FIG. 6, the A-B-S method is realized by determining the optimum index in the noise codebook 12 in addition to the optimum lag determination. The noise codebook 12 stores a predetermined number (for example, 128) of noise signals of several tens of samples (40 samples = 1 vector in the above example) different from each other. The vector thus stored in the noise codebook 12 is called a noise code vector. When the number of noise code vectors stored in the noise code book 12 is 128, each noise code vector can be specified by the 7-bit index. The content of the noise codebook 12 is the same on the encoding side and the decoding side. In this way, the information about the noise codebook 12 can be given only by transmitting the information of the optimum index from the encoding side to the decoding side.

The error power minimization unit 26 performs the above-mentioned optimum lag search for the excitation codebook 10 and then optimizes the index for the noise codebook 12.

The optimum lag and the optimum index thus obtained are transmitted from the encoding side to the decoding side. That is, the output of the device on the encoding side shown in FIG. 6 is the optimum lag and the optimum index obtained by the error power minimization unit 26. In other words, in the CELP coding method, the input speech is represented by the output that drives the weighting synthesis filter 20 with the sum of the excitation code vector and the noise code vector. Quantization in such vector units is called vector quantization.

In addition to this, information as a driving condition of the weighting synthesis filter 20, that is, gains β and γ multiplied by each code vector and various parameters of the weighting synthesis filter 20 are also transmitted. Gains β and γ transmitted
Are optimum gains optimized using the error powers. For example, the optimum value γ _opt of the gain γ is γ = γ _opt that satisfies the condition that the error power E _i obtained by the following equation is dE _i / dγ = 0.

E _i = Σ {p (n) -γ × e _i (n) * h
(N−t)} ² where E _i is the error power of the i th noise code vector, e _i (n) is the n th sample in the i th noise code vector, and p (n) has already been calculated. A signal obtained by subtracting a signal obtained by weighting and combining the optimal sound source vector from the perceptually weighted input signal, and h (n−t) is an impulse response of the weighting synthesis filter 20. * Indicates convolution, and Σ indicates the sum total of 40 samples. Optimal gain γ for some i
_{By obtaining opt} and substituting this value into the above equation, the true error power E _i (opt) is obtained for each i. As is apparent from this, the optimum gain γ _opt also differs when i is different. The same applies to the optimum value β _opt of the gain β.

On the decoding side, the same codebook as on the encoding side is used to obtain each code vector based on the transmitted information, drive the synthesizing filter, and output the reproduced signal.

The frame length or subframe length is
It is fixed regardless of the period of the sound source signal. When the optimum lag search range is set to the range of −20 to −146 samples as described above, if the lag is 20 samples, for example, 40 samples cannot be taken starting from this lag. In such a case, the error power minimization unit 26 generates the 40 samples by repeating the latest 20 samples twice, and uses this for the linear addition and the weighted combination. Further, when a voice with a high pitch frequency such as a female voice is input, the optimum lag becomes short. In such a case, a lag in which the length of the sound source signal is less than the frame length may be optimal. In this case, the sound source signal is repeatedly used and used as the sound source signal.

[0018]

In the region where the level of the input voice is steady, the above-described processing can be suitably performed for encoding, but in the region of the beginning and end of the word, the region is Even if the waveforms are the same, the levels are different. That is, since the input voice has rises or falls in the beginning and end regions, it is considered that rises or falls are generated even if generated by the sound source codebook. If this area is treated in the same way as a stationary area, quantization noise of the reproduced signal reproduced by the decoder due to the information transmitted from the encoding side to the decoding side especially occurs at the beginning and end of the word, and the reproduced voice quality deteriorates. May occur. As a countermeasure against this, a method of shortening the frame length or the subframe length can be considered, but in this case, the number of bits per unit time increases and the coding efficiency becomes poor.

The present invention has been made to solve the above problems, and when the input speech having a short pitch period such as a female voice is coded by the CELP coding method, the beginning and The purpose is to improve the reproduced voice quality of the ending.

[0020]

In order to achieve such an object, the present invention provides a source code vector obtained by an source codebook having a predetermined frame or subframe length and a noise obtained by a noise codebook. An optimal lag and a search are performed in the source codebook so that the error power obtained by obtaining the linear combination of the code vectors and inputting it to the weighting synthesis filter and comparing the output signal with the perceptually weighted input speech is minimized. Generates a sound source code vector by setting the optimum gain, transmitting information including the set optimum lag and optimum gain, and repeating the sound source waveform related to the optimum lag when the optimum lag is shorter than the frame or subframe length. In a CELP coding method using the A-b-S method, whether a certain frame or subframe When the frame or subframe power of the input speech rises or falls over the next frame or subframe, the amplitude of the sound source waveform included in the signal output from the sound source codebook is set to the rise or rise of the input speech. It is characterized in that a transition is made in the same manner as the amplitude transition associated with the falling.

[0021]

In the present invention, the signal output from the sound source codebook is used when the frame or subframe power of the input speech rises or falls from one frame or subframe to the next frame or subframe. The amplitude of the included sound source waveform makes a transition similar to the amplitude transition associated with the rising or falling of the input voice.
Therefore, even when encoding an input voice having a short pitch period such as a female voice, since the slope relating to the rising or falling of the input voice is weighted to the sound source waveform, a proper sound source waveform can be obtained and a proper pitch can be extracted. You can do it. Further, when the lag shorter than the frame or subframe length is optimal, the quantization noise of the frame or subframe relating to the rising and falling can be reduced, so that the quantification of the frame or subframe relating to the beginning and ending of the input speech is performed. Noise is reduced and the reproduced voice quality is improved.

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT A preferred embodiment of the present invention will be described below with reference to the drawings. It should be noted that the same components as those of the conventional example shown in FIG.

FIG. 1 is a block diagram showing a method according to an embodiment of the present invention. The device shown in this figure is more than the conventional device shown in FIG.
This is a configuration in which a fall detection circuit 28 and a sound source waveform generation circuit 30 are added. Further, the output of the sound source codebook 10 is not directly input to the coefficient multiplier 14, but is input to the coefficient multiplier 14 via the sound source waveform generation circuit 30.

The rising / falling detection circuit 28 is
The frame (or subframe) power of the input voice is calculated, and the rising and falling edges of the input voice are detected by discriminating the threshold value of this power, and the slopes of the rising and falling edges are obtained. Sound source waveform generation circuit 30
When the rising / falling detection circuit 28 detects the frame (or subframe) power of the input voice and rising / falling, the amplitude of the sound source code vector is linear according to the slope of the rising / falling. Change.

Now, consider a case where the frame (or sub-frame) power of the input voice rises. The input voice is silent up to the previous frame (or subframe), becomes voiced from the current frame (or subframe), that is, the corresponding frame (or subframe), and the power of the next frame (or subframe) As shown in 2, it is assumed that the number has increased compared to the corresponding frame (or subframe).

The rising / falling detection circuit 28 is
This rising is detected by the threshold judgment,
The rising slope is detected as shown in FIG. The sound source waveform generation circuit 30 follows the inclinations of FIGS.
As shown in, the output amplitude is changed within the corresponding frame (or subframe). The level of the signal output from the sound source codebook 10 is constant in the same frame (or subframe), and the gain β of the coefficient multiplier 14 is optimally set by the error power minimization unit 26.

Therefore, by such processing, even when the input voice rises or falls at a lag shorter than the frame length, the inclination is weighted to the sound source waveform, so that a proper sound source waveform is obtained and a proper pitch is obtained. Can be extracted. Further, when the lag shorter than the frame length is optimal, the quantization noise of the frame relating to the rising and falling can be reduced, so that the generation of the quantization noise at the beginning and the ending of the input voice can be prevented, and the reproduced voice quality can be improved. Can be improved. Furthermore, in the CELP coding method, since the frame power is also normally transmitted as a transmission parameter, no new transmission parameter is required to implement this embodiment.

In this example, the lag is 1 of the frame length.
/ 2. Unlike this example, when the source waveform of the lag being searched for does not fit in an integral number in one frame, the level of the last repetitive waveform (not including one waveform) is equal to the level of the previous source waveform. To do.

[0029]

As described above, according to the present invention,
When the frame or subframe power of the input voice rises or falls from one frame or subframe to the next frame or subframe, the amplitude of the sound source waveform included in the signal output from the sound source codebook, Since the transition is performed in the same manner as the amplitude transition related to the rising or falling of the input voice, even when the input voice having a short pitch period such as a female voice is encoded, the slope related to the rising or the fall thereof is the source waveform. , A proper sound source waveform is obtained, and a proper pitch can be extracted. Further, when the lag shorter than the frame or subframe length is optimal, the quantization noise of the frame or subframe relating to the rising and falling can be reduced, so that the quantification of the frame or subframe relating to the beginning and ending of the input speech is performed. Noise is reduced and the reproduced voice quality is improved. Also, no new transmission parameters are required when transmitting frame or subframe power to the decoding side.

[Brief description of drawings]

FIG. 1 is a block diagram showing a method according to an embodiment of the present invention as a block configuration of an apparatus.

FIG. 2 is a diagram showing a transition of frame power from a corresponding frame to a next frame at the time of rising.

FIG. 3 is a diagram showing an amplitude transition associated with this power transition.

FIG. 4 is a diagram showing a slope of an amplitude level given to a sound source waveform in a corresponding frame.

FIG. 5 is a diagram showing a sound source waveform related to a corresponding frame.

FIG. 6 is a block diagram showing a method according to a conventional example as a block configuration of an apparatus.

[Explanation of symbols]

 10 Sound Source Codebook 12 Noise Codebook 14,16 Coefficient Multiplier 18 Adder 20 Weighting Synthesis Filter 22 Auditory Weighting Filter 24 Subtractor 26 Error Power Minimization Unit 28 Rise / Fall Detection Circuit 30 Sound Source Waveform Generation Circuit

Claims (1)

[Claims] 1. A linear combination of an excitation code vector having a predetermined frame or sub-frame length and obtained by an excitation codebook and a noise code vector obtained by a noise codebook is obtained, input to a weighting synthesis filter, and its output signal is obtained. The optimum lag and the optimum gain are set in the search of the sound source codebook so that the error power obtained by comparing the input sound weighted with the perceptual weight is minimized, and information including the set optimum lag and the optimum gain is set. In the CELP coding method using the ABS method, in which the excitation code vector is generated by repeating the excitation waveform related to the optimal lag when the optimal lag is shorter than the frame or subframe length while transmitting, Or the frame of the input audio from the subframe to the next frame or subframe Is characterized in that when the subframe power is rising or falling, the amplitude of the sound source waveform included in the signal output from the sound source codebook is transited in the same manner as the amplitude transition related to the rising or falling of the input voice. Source sound source waveform generation method.