JPH10222194A - Discriminating method for voice sound and voiceless sound in voice coding - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for discriminating surely a voice sound and a voiceless sound in voice coding. SOLUTION: Voice frame data of an input voice is divided into four sub- frames, further it is discriminated whether an input voice in each sub-frame is a voice sound or a voiceless sound as the following. NC values (normalized correlation value) of each sub-frame are compared with a high critical value and a low critical value respectively, magnitude of an energy quantity value and a LSP (line spectrum pair of element) coefficient value of the sub-frame is discriminated respectively by a discrimination step discriminating whether it is stable or unstable, when both are larger than a set critical value, a discrimination step for an energy ratio of low frequency band/high frequency band (LOH) is performed, it is discriminated whether each sub-frame is in a critical value or more or not in a LOH value discriminating step, when it is 'yes', a sub-frame is discriminated as a voice sound signal and when it is 'no', a sub-frame is discriminated as a voiceless signal.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a kind of speech coding technique, and more particularly to a method used in speech coding technique to discriminate between voiced sound and unvoiced sound.

[0002]

2. Description of the Related Art In the technology of speech synthesis, linear predictive coding (Linear Predictive Coding) is used.
ding (LPC) is generally used. In this LPC method, the LPC-10 audio encoder uses
Widely used for low bit rate audio compression. One L
For PC audio encoders, it has been an important issue how to accurately identify whether an input audio signal is voiced or unvoiced. This is because the voiced / unvoiced sound discrimination process can greatly affect the output quality of speech synthesis.

[0003] Fig. 1 is a block diagram of a traditional speech coding technique. In the block in the figure, an impulse train generator 11 (Impulse Train)
nGenerator), random noise generator 12 (Random Noise Generator)
r), voiced / unvoiced sound selector switch 13 (voice
d / unvoiced Switch), gain unit 14 (Gain Unit), LPC filter 15 (L
inner Predictive Coding Fi
lter), LPC filter control variable setting unit 16
Is included.

[0004] Periodic impulse trains (Periodic Impulses) generated by the impulse train generator 11
e Train) or random noise generator 1
2 generated noise signal (White Noise)
Is appropriately switched by the voice / unvoiced switch 13 according to the type attribute of the input signal.
The gain of the signal is performed via the gain unit 14, so that the level of the signal is adjusted. Further, the LPC filter 15 sets the LPC variable (LPC Parameters) set in the LPC filter control variable setting unit 16.
Filter, and finally, the LPC filter 15
, An audio output S (n) is performed from the output terminal.

[0005] When performing the above-described speech identification step, the identification device will use the speech frame (Sp) of each one of the input speeches.
ech Frame), its voiced / unvoiced sound discrimination method, pitch period (Pitch Period), LP
Update the C variable and the gain value (GainValue). The purpose is to be able to follow the changing state of the input voice. In the current typical technique, each one audio frame covers 160 samples, ie
Samples are taken every 0.02 seconds within the size of one predetermined audio frame.

[0006] In the above-mentioned speech recognition, a method of discriminating a voiced sound or an unvoiced sound among them is traditionally a method based on the strength of a correlation pitch. For example, if the normalized cross-correlation value (Normalized C
Ross Correlation Value; NC
Is larger than a preset critical value, for example, 0.4 or more, the speech frame is determined to be a normal speech signal. At this time, the speech synthesizer excites the LPC filter with a periodic pulse train. Conversely, if the NC value is the critical value 0.4
If smaller, the speech frame is determined to belong to the unvoiced sound signal, and the speech synthesizer excites the LPC filter with a random noise generator. The above-described definition of the NC value is based on Equation 2 below.

(Equation 2)

However, for an unstable voice signal, that is, one that fluctuates in an uncertain level area above and below the critical value, the NC value below the critical value of 0.4 can be very small. However, the simple determination method described above cannot accurately determine whether the signal is a voiced sound signal or an unvoiced sound signal. Therefore, at the time of actual application, there is a possibility that a mistake is generated.

In order to overcome the above problem and increase the accuracy of the discrimination, in the known technique, in addition to the above-described discrimination of the NC value, it is necessary to further execute the discrimination of the amount of energy of the voice signal. Has achieved the purpose of making relatively accurate discrimination.

[0009] In the known technology, there is another type of voiced / unvoiced sound discrimination method. When determining the energy of the audio signal based on this other known technique, the following two situations are included. a. Generally speaking, the voice energy of unvoiced sounds is lower than that of voiced sounds, and the root mean square value (RM)
S) is obtained by the following Expression 3.

(Equation 3) Among them, N represents the entire audio frame of the input audio signal. b. Zero-crossing rate (ZC) The definition is defined as the number of zero-crossings in the entire voice frame, and is calculated by the following equation (4).

(Equation 4) In the above-described speech coding technique, each one speech frame contains 160 samples, and in terms of the number of bits, each speech frame is composed of a 34-bit LPC variable, a 6-bit pitch, 1-bit voiced / unvoiced sound,
Includes a total of 48 bits of 7-bit gain values.

As described above, in encoding speech, it is important to accurately determine whether the input speech signal is voiced or unvoiced. The discrimination process greatly affects the output quality of speech synthesis. If the unvoiced sound is mistaken for a voiced sound in the voiced / unvoiced sound discrimination process, the output synthesized voice will be a sound like a groan, and if the voiced sound is mistaken for an unvoiced sound, the output will be made. The synthesized speech sounds like an elaborate sound. The traditional techniques described above have not been able to effectively solve this problem.

Further, in the above-mentioned second type of traditional technology,
One bit determines the state of voiced or unvoiced sound in a voice frame, and thus implies the critical state between voiced and unvoiced sounds. For this reason, since the entire speech frame was in the critical area and it was not determined whether the speech sound was voiced or unvoiced, the output synthesized speech often seemed to have noise.

[0012]

The shortcomings of the known techniques described above indicate that traditional speech coding techniques need improvement. Accordingly, it is a primary object of the present invention to provide a kind of improved technique for speech coding, thereby providing excellent speech synthesis output quality during speech coding.

Another object of the present invention is to provide a method for accurately discriminating between voiced sound and unvoiced sound, which is used during speech coding. An object of the present invention is to make it possible to distinguish between a vocal sound and an unvoiced sound.

Still another object of the present invention is to provide a method for discriminating voiced / unvoiced sounds of a kind of quadrant.
Voiced / Unvoiced Decision
(Scheme), which divides each one audio frame in the input audio signal into four subframes, and then, for each one subframe,
On the basis of the correlation variables, it is determined whether the sub-frame is voiced or unvoiced sound, and based on the result of the determination, an accurate and natural sound signal is output at the speech synthesis output terminal.

Still another object of the present invention is to provide a method for accurately discriminating voiced / unvoiced sounds in a voice frame of an input voice signal. In the steps of the present invention, first, the voice frame of the input voice is divided into four sub-frames, and the NC values (normalized cross-correlation values) of the four sub-frames are sequentially higher than a high critical value (for example, 0.7). Is determined. Thereafter, it is further determined whether the NC value is smaller than a low critical value (for example, 0.4). After the above two discriminating steps, signals that clearly belong to voiced and unvoiced sounds are discriminated, and subsequently, signals that intervene between the aforementioned signals that clearly belong to voiced and unvoiced sounds are discriminated. That is, if it is determined in the above steps that the NC value is not smaller than the low critical value, a stable / unstable determining step is performed, and the energy amount value of the subframe and the line spectrum pair (Line Spectrum Pair) are executed. ;
LSP) coefficient value, that is, the magnitude of the LSP coefficient value is determined, and if the energy amount value and the LSP coefficient value are not greater than a predetermined critical value, it is determined that the audio signal exhibits a stable state, and All the attributes of the subframe are set to be the same as the voiced / unvoiced state of the last one subframe in the previous one voice frame, and if the energy amount value and the LSP coefficient value are set to the critical value during the above-mentioned steps, If it is determined to be larger, the energy ratio of the low frequency band to the high frequency band (Low to
High Band Energy Ratio Va
l; LOH), and determines whether the LOH value of each subframe is greater than a threshold value. If the LOH value is greater than the threshold value, the subframe is determined to be a voiced speech signal. If not, the subframe is determined to be an unvoiced speech signal. Similarly, the determination proceeds for the next one subframe, and thus the determination is performed for all four subframes.

[0016]

The invention of claim 1 is a method for discriminating voiced and unvoiced sounds in a kind of speech coding, which is used for discriminating attributes of speech frame data of input speech. The method comprises the following steps a to f: a. Split the audio frame data of the current input audio into four subframes, b. Normalized cross-correlation values of the four subframes (Norm
aligned Cross Correlation
Value), that is, whether the NC value is greater than or equal to one high critical value. If the result of the determination is yes, all four subframes in the input current voice frame are voiced sound signals. Is determined to be c. If the NC value of the subframe is not equal to or higher than the high critical value in step b, it is determined whether the NC value is smaller than one low critical value. Determining that each of the two subframes belongs to an unvoiced sound signal; d. If it is determined in step c that the NC value is not smaller than the low critical value, a step of determining whether the NC value is stable or unstable is performed, and the energy value of the subframe and the line spectrum pair are used.
Determining the magnitude of the coefficient value, ie, the magnitude of the LSP coefficient value, e. If the energy value and the LSP coefficient value are not greater than the set threshold values, it is determined that the audio signal is in a stable state, and all attributes of the four subframes are changed to the last one in the previous audio frame. Set the same as the voiced or unvoiced state of the subframe; f. In step e above, if the energy value and L
If the SP coefficient value is greater than the set threshold value, the low-to-high band energy ratio of the subframe is low to high band energy.
Ratio Value), i.e., a step of determining the LOH value is performed, and it is determined whether each LOH value is greater than a certain threshold value. If the LOH value is greater than the threshold value, the subframe is determined to be a voiced sound signal. If it is not large, the subframe is determined as an unvoiced sound signal, the determination proceeds to the next one subframe, and the determination is completed after finishing the determination for all four subframes. This is a method for discriminating voiced and unvoiced sounds in speech coding.

According to a second aspect of the present invention, the high critical value used to determine the NC value of the subframe in step b is set to 0.7. For discriminating voiced and unvoiced sounds.

According to a third aspect of the present invention, the low critical value used when determining the NC value of the subframe in step c is set to 0.4. For discriminating voiced and unvoiced sounds.

According to a fourth aspect of the present invention, in the step of determining whether the step d is stable or unstable, in the determination of the energy amount of the subframe, the difference value between the previous one energy amount and the current energy amount is set. 2. A method for discriminating voiced and unvoiced sounds in speech coding according to claim 1, wherein it is determined whether or not the threshold value is equal to or greater than a certain critical value.

According to a fifth aspect of the present invention, there is provided the method for discriminating voiced and unvoiced sounds in speech coding according to the fourth aspect, wherein the set critical value is set to 0.45 in the energy amount determining step.

According to a sixth aspect of the present invention, in the step of determining whether the step d is stable or unstable, in determining the LSP coefficient value of the subframe, the previous average LSP coefficient value and the current LSP coefficient value are determined.
The method according to claim 1, wherein a difference value between the SP coefficient value and the SP coefficient value is determined.

According to the seventh aspect of the present invention, the LSP
The method according to claim 6, wherein the critical value used in the coefficient value discriminating step is set to 0.4.

According to an eighth aspect of the present invention, in the step f, during the step of determining the LOH value of the subframe, the definition of the LOH is expressed by the following equation 1.

## EQU1 ## where i represents the ith subframe, and S
₂ 1p1k represents the signal obtained after the original signal has passed through the 1k low-pass filter. In the definition, the energy amount ratio of the signal lower than 1 KHz and the signal higher than 1 KHz in the audio signal is one window length W. The definition of the so-called window length W is that the pitch is Nsubframe
If greater than, W = pitch (pitch) Nsubframe / 2 or more and Nsubfra
If less than me, then W = 2 * pitch, where Nsubframe indicates the subframe length of the sample, and in the definition of LOH, the silence threshold Tsil is the maximum speech value of the current speech frame; Value is 1K
Hz high pass filter may be added into the energy content of the audio signal which has passed through the high pass filter, thereby giving low energy content voiced signals a tendency to be selected as unvoiced.
t (j) is the center position of each one subframe, and its definition is: doffset (j) = Nsubframe * (j−1)
/ 2), j = 1 to 4, wherein j represents the number of a subframe, and wherein j is the number of a subframe, and wherein the method of discriminating voiced and unvoiced sounds in speech coding according to claim 1 is provided.

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In a discrimination method according to the present invention, an audio frame of an input audio signal is divided into four subframes (Subfr).
ame), and then, based on variables correlated to each one of the sub-frames, it is substantially determined whether each one of the sub-frames is voiced or unvoiced. The above variables are
NC, energy amount, line spectrum vs. even coefficient (line
Spectrum Pair (LSP), and a low frequency band to high frequency band energy ratio (Low to
High Band Energy Ratio Va
lue; LOH).

The following is the determination step of the present invention. FIG.
Is a determination flowchart according to the present invention.
The steps include: After the start step 101 of the flowchart, step 102 is executed. In step 102, the current audio frame data is obtained.
Subsequently, step 103 for determining whether the NC value is equal to or higher than one high critical value 0.7 is executed. See the above description for the definition of the NC value. If the determination is yes, step 104 is executed. In step 104, it is determined whether all of the four sub-frames in the input current audio frame data are voiced sound signals, and then the determination process is terminated.

If it is determined in step 102 that the NC value is not higher than the high critical value 0.7, it is determined in step 105 whether the NC value is lower than the low critical value 0.4. It is determined, and if yes, it is determined that all four sub-frames in the audio frame belong to the unvoiced sound signal, and thereafter, the determination process ends.

After the above-described steps 102 and 103, signals that clearly belong to voiced or unvoiced sounds are determined. Subsequently, signals intervening between the signals belonging to the above-mentioned apparently voiced or unvoiced sounds are discriminated, and a single step 102, a step 1
It is impossible to make an accurate determination of voiced / unvoiced sound by the NC value determination step 03, and therefore the problem of the present invention can be solved by the following determination method. Therefore, it can be said that the following determination steps are key steps that show the features of the present invention.

If it is determined in step 105 that the NC value is not smaller than 0.4, it is determined whether the NC value is stable or unstable (Stationary / nonst).
ationary Decision; S / NS D
execution). In this one step, two discrimination items are included, one of which is the discrimination of the energy amount, which is one energy amount (P
difference between the current energy (current energy) and the current energy amount (current energy), that is, dis
(PrEng, CuEng). In order to further increase the accuracy of the S / NS determination, LSP coefficient determination is further included in this one step.
This LSP coefficient is obtained from the LPC equalizer. This LS
In the determination of the P coefficient, the previous one average LSP (Past
average LSP) and current LSP (Curren)
t LSP), ie, dis (PaLSP, CuL)
SP). During the S / NS determination step of step 107, a. dis (PrEng, CuEng) is 0.45 or more, and b. determine if dis (PaLSP, CuLSP) is greater than or equal to 0.4; if the result is no, indicate that the audio signal is in a stable state, execute step 108, and All attributes are set to be the same as the voiced or unvoiced state of the last one subframe in the previous one voice frame. Conversely, if the result of the determination of the difference value in step 107 is yes (ie, indicating that the change in energy amount or LSP coefficient is extremely fast), the LOH determination step (steps 109 to 113) is executed. , For each subframe, to determine whether it is voiced or unvoiced, thereby obtaining an accurate determination result. The definition of the so-called LOH determination is as shown in Equation 1 below.

## EQU1 ## where i represents the ith subframe, and S
₂ 1p1k represents the signal obtained after the original signal has passed through the 1k low-pass filter. 1KHz in audio signal during definition
The energy ratio between the lower one and the one higher than 1 KHz is mutually offset by one window length W, and the definition of the so-called window length W is as follows. W = pitch (pitch) Pitch is Nsubfra
When the pitch is larger than me W = 2 * pitch When the pitch is equal to or larger than Nsubframe / 2 and smaller than Nsubframe, Nsubframe indicates the subframe length of the sample. In addition, during the definition of LOH, one silence critical value Tsil is selected as the maximum audio value of the current audio frame, and the Tsil value can be added to the energy amount of the audio signal passed through the 1 KHz high-pass filter. , Thereby tending to select voiced signals of low energy content as unvoiced sounds. doffset (j) is the center position of each one subframe, and its definition is: doffset (j) = Nsubframe * (j−1)
/ 2), j = 1 to 4, j indicates the number of the subframe. In the LOH determination flowchart of the present invention, it is first determined in step 110 whether or not the LOH of the first subframe (see the above definition) is greater than 1; Step 112 is executed to determine that the subframe is a voiced sound signal. If no, step 111 is executed, that is, it is determined that the subframe is an unvoiced signal. Then, step 113
And 119, the determination proceeds for the next one subframe, and the determination ends for all four subframes. That is, after the above-described LOH determination, if the LOH value of each one sub-frame is larger than one critical value, the sub-frame is determined to be voiced; if not, the sub-frame is determined to be unvoiced. After the determination of all four sub-frames of one voice frame is completed, the process proceeds to encoding based on the result. In the present invention, four sub-frames are coded with only 3 bits, as shown in FIG. Among them, 1 indicates a voiced sound and 0 indicates an unvoiced sound.

After obtaining the index values shown in FIG. 3, the corresponding values are stored and the encoding process is completed. Then, in an actual application, decoding is performed using a well-known speech synthesis technique, and the necessary synthesized speech is decoded. Occurs.

[0030]

The present invention provides a kind of improved speech coding technology, whereby an excellent speech synthesis output quality can be obtained in the speech coding process. The present invention further provides a method for accurately distinguishing voiced or unvoiced sounds used during speech coding, whereby the method for accurately determining whether a speech frame in an input speech signal is voiced or unvoiced. It becomes possible to determine. The present invention further relates to a method of discriminating voiced / unvoiced sounds of a quadrant (Quarter Voice).
d / Unvoiced Decision Schem
e).

[Brief description of the drawings]

FIG. 1 is a basic block diagram of a traditional speech coding technique.

FIG. 2 is a determination flowchart of the present invention.

FIG. 3 is a code table in which four subframes are encoded with 3 bits in the present invention.

[Explanation of symbols]

 Reference Signs List 11 impulse train generator 12 random noise generator 13 voiced / unvoiced sound changeover switch 14 gain unit 15 LPC filter 16 LPC filter control variable setting unit

Claims (8)

[Claims] 1. A method of discriminating voiced and unvoiced sounds in a type of speech encoding, which is used to identify attributes of speech frame data of an input speech. Encompassing the steps: a. Split the audio frame data of the current input audio into four subframes, b. Normalized cross-correlation values of the four subframes (Norm
aligned Cross Correlation
Value), that is, whether the NC value is greater than or equal to one high critical value. If the result of the determination is yes, all four subframes in the input current voice frame are voiced sound signals. Is determined to be c. If the NC value of the subframe is not equal to or higher than the high critical value in step b, it is determined whether the NC value is smaller than one low critical value. Determining that each of the two subframes belongs to an unvoiced sound signal; d. If it is determined in step c that the NC value is not smaller than the low critical value, a step of determining whether the NC value is stable or unstable is performed, and the energy value of the subframe and the line spectrum pair are used.
Determining the magnitude of the coefficient value, ie, the magnitude of the LSP coefficient value, e. If the energy value and the LSP coefficient value are not greater than the set threshold values, it is determined that the audio signal is in a stable state, and all attributes of the four subframes are changed to the last one in the previous audio frame. Set the same as the voiced or unvoiced state of the subframe; f. In step e above, if the energy value and L
If the SP coefficient value is greater than the set threshold value, the low-to-high band energy ratio of the subframe is low to high band energy.
Ratio Value), i.e., a step of determining the LOH value is performed, and it is determined whether each LOH value is greater than a certain threshold value. If the LOH value is greater than the threshold value, the subframe is determined to be a voiced sound signal. If it is not large, the subframe is determined to be a signal of unvoiced sound, the determination proceeds to the next one subframe, and the determination is completed after finishing the determination for all four subframes. A method for distinguishing voiced and unvoiced sounds in speech coding. 2. The voiced speech and unvoiced speech in speech coding according to claim 1, wherein the high critical value used when discriminating the NC value of the subframe in step b is set to 0.7. Identification method. 3. The voiced and unvoiced speech in speech coding according to claim 1, wherein the low critical value used in determining the NC value of the subframe in step c is set to 0.4. Identification method. 4. In the step of determining whether the step d is stable or unstable, in the determination of the energy amount value of the subframe,
Determine whether the difference between the previous one energy amount and the current energy amount is equal to or more than a set critical value,
A method for discriminating voiced and unvoiced sounds in speech coding according to claim 1. 5. The method for discriminating voiced and unvoiced sounds in speech coding according to claim 4, wherein in the step of determining the energy amount value, the set critical value is 0.45. 6. In the step of determining whether the step d is stable or unstable, in determining the LSP coefficient value of the subframe, a difference value between a previous average LSP coefficient value and a current LSP coefficient value is determined. The method for distinguishing voiced and unvoiced sounds in speech coding according to claim 1, characterized in that: 7. The method according to claim 6, wherein the threshold value used in the step of determining the LSP coefficient value of the subframe is set to 0.4. 8. In the step f, the LOH of the subframe is
During the value determination step, LOH is defined by the following equation (1). Where i represents the i-th subframe, S ₂ 1p1k represents the signal obtained after the original signal has passed through the 1k low-pass filter, and, during definition, is lower than 1 KHz in the audio signal. The energy amount ratio of higher than 1 KHz is divided by one window length W. The definition of the window length W is as follows: When the pitch is larger than Nsubframe, W = pitch (pitch) When the pitch is Nsubframe / 2 or more, Nsubframe
If less than me, then W = 2 * pitch, where Nsubframe indicates the subframe length of the sample, and in the definition of LOH, the silence threshold Tsil is the maximum speech value of the current speech frame; Value is 1K
Hz high pass filter may be added into the energy content of the audio signal which has passed through the high pass filter, thereby giving low energy content voiced signals a tendency to be selected as unvoiced.
t (j) is the center position of each one subframe, and its definition is: doffset (j) = Nsubframe * (j−1)
/ 2), j = 1 to 4, wherein j represents the number of a subframe, wherein the voiced sound and the unvoiced sound are distinguished from each other in voice coding according to claim 1.