JPH06348293A - Voice information analyzing device - Google Patents

Abstract

PURPOSE:To effectively eliminate noise, whose features change dynamically, from a voice information analyzing device. CONSTITUTION:When a sound and silence detection section 720 discriminates it as no sound, a noise processing section 690 obtains noise features from the power spectrum of the input voice outputted by an axis transformation section 640 and stores them in a noise table 710. A no sound frame setting section 700 outputs the data, which signify silence, as a normalized waveform series 682. When it is discriminated as a sound, a noise eliminating section 660 eliminates noise from the power spectrum column outputted by the section 640 using the noise features stored in the table 710. An inverse FFT section 670 performs an inverse FFT of the power spectrum from which noise is eliminated. A normalizing section 680 normalizes the output of the section 670 using the pitch information received from a pitch extracting section 650 and outputs them as a normalized waveform series 682. Thus, by constantly extracting and eliminating environmental noise features that are mixed in the input signals and are constantly changing, noise elimination is securely performed under any circumstances.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice information analysis apparatus for analyzing voice information, and more particularly to a technique for removing a dynamically changing noise component from a voice signal.

[0002]

2. Description of the Related Art As a conventional technique for removing noise from an input audio signal, a technique described in JP-A-2-278298 and a technique described in JP-A-1-75593 are known.

The technique described in Japanese Patent Laid-Open No. 2-278298 is a technique in which a filter is subdivided and used to fixedly remove a specific frequency, and its sound quality depends on the frequency to be removed. Further, the above-mentioned JP-A-1-7559.
The technique described in Japanese Patent Publication No. 3 uses a neural network for noise removal to previously learn the features of noise /
This is a technique for extracting and removing noise using this. The sound quality depends on the noise selection that mixes the characteristics of the noise with the data (voice + noise) used for learning / extraction.

[0004]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
According to the technique of removing a specific frequency by using the noise removal filter described in Japanese Patent No. 8298, since the frequency to be removed is fixed, it dynamically changes with the movement and change of the speaker and the background. It is not suitable for removing noise with changing characteristics.

Even in the technique of extracting / learning the characteristics of noise described in Japanese Patent Laid-Open No. 1-75559, it takes a long time to learn noise due to the nature of the neural network. It is not suitable for removing noise whose characteristics change dynamically as the background moves or changes.

Further, when performing voice information analysis, in order to remove noise from the voice signal of the voice information analysis target,
When these techniques are applied, in any case, the noise removal process must be performed separately from the voice analysis process as a pre-process of the voice information analysis process. For this reason, when performing real-time voice information analysis processing, there is a possibility that it will be overloaded, and it is realized by limiting the function of voice information analysis, or by executing each processing by a different processor. There may be a need.

Therefore, an object of the present invention is to provide a voice information analysis apparatus which can efficiently remove noise whose characteristics change dynamically from a voice signal which is a target of voice information analysis.

[0008]

[Means for Solving the Problems] To achieve the above object,
The present invention is a voice information analysis method for outputting an analysis result obtained by analyzing voices represented by frame data obtained by collecting voice sampling data for a certain period of time, and voices other than noise are included in voices represented by each frame data. If it is determined that a voice other than noise is not included, the feature of noise included in the voice represented by the frame data is extracted from the frame data and stored, As the analysis result of the voice indicated by the frame data, a step of executing a silent system process of outputting information representing the analysis result of a silence voice prepared in advance, and when it is determined that a voice other than noise is included , The noise feature stored in the previous silence processing was removed from the voice represented by the frame data, and the voice represented by the frame data from which the noise feature was removed was analyzed. Having and executing the sound system processing of outputting the result of analysis to provide voice information analyzing method comprising.

[0009]

According to the voice information analysis method of the present invention, it is determined whether or not the voice represented by each frame data includes voice other than noise, and it is determined that voice other than noise is not included. In this case, the characteristic of noise included in the voice represented by the frame data is extracted from the frame data and stored, and the analysis result of the silent voice prepared in advance is used as the analysis result of the voice indicated by the frame data. Executes silent processing that outputs the information that is represented. On the other hand, when it is determined that the sound other than noise is included,
The noise feature stored in the previous silence processing is removed from the voice represented by the frame data, the voice represented by the frame data from which the noise feature is removed is analyzed, and the voiced processing that outputs the analysis result is executed. To do.

Therefore, the latest noise feature extraction is performed during a period in which no voice other than noise is included, that is, in a period in which it is considered that the voice is silent and the analysis result of the silence obtained in advance is sufficient. In the period in which the sound other than the above is included, noise can be removed by using the extracted latest noise feature. Further, since the voiced system processing and the silent system processing do not occur at the same time, the execution load of this processing is small and it can be realized on a single processor without restricting the voice information analysis function.

[0011]

EXAMPLE An example of the present invention will be described below.

First, the first embodiment will be described.

FIG. 5 shows the configuration of a communication system to which the voice information analyzing apparatus according to the present invention is applied.

In the figure, 2000 is a transmitter and 1000 is a receiver.

The transmitting device 2000 transmits to the receiving device 1000 the level information, the quantized data and the pitch information which are obtained by compressing and encoding the voice signal by a method using the voice analysis. The receiving device 1000 decodes the voice from the received information and outputs it.

Here, the transmitter 200 includes a transmitter 9
00, vector quantization unit 800, voice information analysis device 10
0, and the receiving device 1000 includes the receiving unit 110.
0, the vector dequantization unit 1100, and the synthesis unit 1200.
, D / A converter 1300, and buffer memory 1500
And an audio output device 1400.

In the transmission device 2000, the voice information analysis device 100 analyzes the input voice that has been input, and outputs the obtained level information 681 and pitch information 651 to the transmission section 900.
Then, the normalized waveform sequence 682 is sent to the vector quantization unit 800. The vector quantizer 800 converts the received normalized waveform sequence 682 into a vector code, and sends the quantized data 801 obtained by the conversion to the transmitter 900. Transmitter 9
00 receives the level information 681, the pitch information 651, and the quantized data 801, and transmits this to the receiving apparatus 1000 via wire / wireless.

On the other hand, in the receiver 1000, the receiver 1000 receives the information transmitted from the transmitter 2000 and outputs level information 681 ', quantized data 801' and pitch information 651 '. Vector inverse quantizer 1
100 inversely quantizes the output quantized data 801 '. The synthesizer 1200 includes a vector dequantizer 1100.
The waveforms 682 ′ output by the above are superposed for each repeating period (pitch information) based on the pitch information 651 ′, thereby synthesizing the waveforms and stored in the buffer memory 1500. The D / A converter 1300 has a buffer memory 40.
The output of 0 is converted to digital / analog (D / A). The audio output device 1400 outputs the audio obtained by the D / A converter 1300.

The details of the voice information analysis apparatus 100 will be described below.

As shown in FIG. 5, the voice information analysis apparatus 100 includes a voice input section 200 which is a voice input means,
A / D converter 30 for converting input input voice into analog / digital (A / D) and converting into voice sampling data
0 and a buffer memory 400 for sequentially storing the voice sampling data. The voice sampling data for every fixed time (10 to 30 milliseconds) stored in the buffer memory 400 is sent to the voice / non-voice determination unit 500 as the voice sampling data (frame data) 401 for a fixed time.
Further, the voice information analysis device 100 includes a voice / non-voice determination unit 500 for determining voice / non-voice, and the voice information analysis device 100.
Is a normalized waveform sequence 68 from the frame data 401.
2. It has level information 681 and an analysis unit 600 that creates pitch information 651.

First, the voiced / non-voiced determination unit 500 is composed of a voice power determination unit 510, as shown in FIG.
The audio power determination unit 510 obtains the total sum (power) of each element of the frame data 401, and this and a threshold value (fixed value).
After that, if it is smaller than the threshold value, it is determined to be silent, and if it is larger than the threshold value, it is determined to be voiced.
Is output.

If a sufficient effect cannot be obtained with the threshold value, the threshold value may be made variable according to the equation (1). In Expression 1, ω is a weight value that satisfies 0 <ω <1.

Next frame data threshold = ω threshold of previous frame data + (1−ω) current silent power value. ． (Equation 1) That is, when it is determined that there is no sound, the power value determined to be soundless and the threshold value used for the determination of the previous frame data are appropriately weighted, and the added value of these is added to the next frame. The threshold value used to determine the system data is calculated.

Further, the sound power judging section 510 may judge the presence / absence of sound as follows. That is, the power of the determination target frame data is compared with the power of the previous frame data, and if the difference is larger than a predetermined value, it is determined that the voiced / silent state has changed from the previous time, and the difference is If it is smaller than the predetermined value, it is determined that the state of voice / silence has not changed. And the previous frame I remembered
Based on the determination result of the frame data, the presence / absence of the frame data to be determined is determined. Alternatively, if the difference is larger than a predetermined value, the power of the frame data to be judged is further compared with the threshold value to judge whether there is sound or no sound.

Next, as shown in FIG. 2, the analysis unit 600 analyzes the FFT (Fast Fo) from the frame data 401.
urier Transform: Fast Fourier Transform)
FFT section 62 for obtaining the frequency characteristic of frame data by
0 and an FFT data setting unit 610 that sets data in the FFT unit 620. In addition, the FFT unit 620
The power spectrum conversion unit 630 that outputs the square value of the absolute value of the complex number obtained by the FFT, that is, the power spectrum, and the axis conversion unit 640 that converts the vertical axis of the power spectrum from the power spectrum axis to the amplitude axis. is doing. In addition, FFT (Fast Fourier Tr
transform: fast Fourier transform) is a technique for reproducing an original waveform from a sampling value of a signal by frequency and amplitude. DFT (Discrete Fourier Tra)
nsform: Discrete Fourier Transform) has been realized at a higher speed, and such signal processing technology by FFT is “Introduction to Signal Processing” Yoshifumi Amamiya / Yukio Sato Chopsticks Company P106-6.3 “Fast Fourier Transform” etc. Are described in detail in.

The analysis unit 600 includes a voice processing system including an FFT unit 730, a pitch extraction unit 650, a noise removal unit 660, an inverse FFT unit 670, and a normalization unit 680, a noise processing unit 690, and a silence. A silence processing system including a data setting unit 700, a noise table 710,
It has a voiced / silent part 720.

The sound / silence section 720 determines the sound / silence determiner 511 output from the sound / silence determining section 500, and in the case of sound, causes the sound processing system to perform processing, and In some cases, the silent processing system is made to perform the processing.

First, the operation of the silence processing system when the sound / silence unit 720 determines that there is no sound will be described.

In this case, the noise processing section 690 uses the output of the axis converting section 640 to determine the characteristics of the noise and the noise table 7.
Store in 10. That is, for example, the frequency characteristic information and the power spectrum sequence subjected to axis conversion by the axis conversion unit are stored in the noise table 710. Silent frame setting section 7
00 outputs, as the normalized waveform sequence 682, the data 682b representing the normalized waveform sequence of silence to the vector quantization unit 800.

Next, the operation of the voice processing system when the voice / silent unit 720 determines that there is voice will be described.

In this case, the FFT section 730 is used to obtain the pitch period from the logarithmic value (cepstrum) of the power spectrum. The pitch extraction unit 650
From the output of the FFT unit 730, the feature (height) of the voice and the repetition period (pitch information) are extracted. Noise removal unit 66
0 uses the characteristics of noise stored in the noise table 710 to remove noise from the power spectrum sequence in the presence of voice output from the axis conversion unit 640. Inverse FFT unit 67
0 is the inverse FFT of the de-noised power spectrum
To do. The normalization unit 680 normalizes the output of the inverse FFT unit 670 using the pitch information received from the pitch extraction unit, sets the maximum value of the results of the inverse FFT to “1”, and represents a waveform within the pitch. The sequence 682a is output as the column-normalized waveform sequence 682. The level information of the output of the inverse FFT unit 670 is output as level information 681.

The noise removal by the noise removal unit 660 is performed as follows, for example. That is, in the noise removing unit 660, a noise mask table storing weightings of 0.0 to 1.0 for each frequency is created in accordance with the stored contents of the noise table 710, and a power spectrum sequence at the time of sound is created. The corresponding weights are multiplied. The weights are sequentially weighted from 0.0 to 1.0 so that the frequency with a large absolute value of the power spectrum (noise power spectrum) stored in the noise table 710 during silence becomes smaller. . That is, the frequency at which noise is noticeable is multiplied by a value of 1.0 or less to shift the power spectrum value of the power spectrum in the presence of voice from the original value toward the decrease direction. , For frequencies where no noise spectrum appears, 1.
The weighting of 0 leaves the power spectrum value at that frequency unchanged. As a result, a noise-free power spectrum sequence can be obtained.

The second embodiment of the present invention will be described below.

The second embodiment is the same as the first embodiment except that the voice /
Only the configurations of the silence determination unit 500 and the analysis unit 600 are different.

Sound / silence determining section 50 according to the second embodiment.
As shown in FIG. 3, 0 is an FFT data setting unit 610 that sets data for FFT processing of frame data.
And an FFT unit 6 which is means for FFT frame data
20 and a power spectrum conversion unit 630 that obtains the sum of squares of the obtained complex numbers, and a voice power determination unit 510 that obtains the sum of the frame data from the frame data, compares it with a threshold value, and outputs a voice / silence determiner 721. And a sound / silence determination unit 720 that determines the sound / sound determination unit.
And the FFT unit 620 for FFT processing of the power spectrum, and the pitch information extracted from the logarithm of the power spectrum and the cepstrum, and the pitch period is not fixed when there is no sound. ~ P59 4.9 Pitch extraction) is used to determine a voiced state and output a voiced / non-voiced discriminator.

Now, the voice power judging section 510, like the voice power judging section 510 according to the first embodiment, judges the presence / absence of sound based on the power of each frame data. However, according to the above-described determination method, since the presence / absence of sound is determined only by the power, the silence state may be erroneously determined as the presence of sound. Therefore, in the second embodiment,
When the voice power determining unit 510 determines that there is sound, the sound / silence determining unit 720 activates the FFT unit 730 and the pitch extracting unit 650, and further performs sound / silence determination using a pitch cycle.

That is, when the voice power determining section 510 determines that there is sound, the FFT section 730 performs FFT processing on the power spectrum. The pitch extraction unit 650 extracts the log information of the power spectrum and the pitch information from the cepstrum from this output, and the pitch period is not fixed when there is no sound ("Digital signal processing" Sadahi Furui P57-
P59 4.9 Pitch extraction)
A silence state is determined and a sound / silence determiner is output. Note that the pitch extraction unit 650 determines the presence / absence of voice, by converting the pitch information on the time axis into a pitch cycle on the frequency axis, and if there is a maximum value for each pitch cycle in the power spectrum, the presence of a sound is detected. If it does not exist, it may be determined to be silent.

Next, the analysis unit 600 according to the second embodiment.
4, as shown in FIG. 4, an axis conversion unit 640 that converts the vertical axis from the power spectrum axis to the amplitude axis, a noise removal unit 660 that removes noise, and an inverse FF that is means for inverse FFT.
A T section 670, a normalization section 680 for setting the maximum value of the result of the inverse FFT to “1”, a voiced / silent section 920,
A noise processing unit 690 for extracting characteristics of noise, and a silence data setting unit 7 for setting data for outputting silence.
00 and a noise table 710. The individual operation of each part is the same as that of the corresponding part of the first embodiment. However, in the second embodiment, the noise removing unit 660 and the inverse FFT are used.
The unit 670 and the normalization unit 680 form a voice processing system. Then, the voiced / silent part 920 makes the pitch extraction part 65
Only when the voiced / non-voiced discriminator output by 0 indicates a voiced voice, this voiced voice processing system is caused to perform processing. On the other hand, the processing of the silence processing system including the noise processing unit 690 and the silence data setting unit 700 is performed by the sound / silence unit 920 and the sound / silence unit 920.
This is performed when at least one of the sound / silence section 720 of the silence determination section 500 determines silence.

As described above, according to the present embodiment, the constantly changing ambient noise is always extracted and the noise feature is removed, so that the noise can be removed accurately in any situation.

By the way, the processing performed by each part of the voiced / non-voiced determination section 500 and the analysis section 600 according to the first and second embodiments can be realized as a program running on a processor. In this case, when there is no sound, only the processing of the sound processing system is performed, and when there is sound, only the processing of the sound processing system needs to be performed, and the preprocessing of the silence processing system and the preprocessing of the sound processing system are common. Therefore, the processing load of the processor is smaller than that of the conventional technique in which noise removal is performed as a pre-process of speech analysis, and the program can be implemented on a single processor.

In the above embodiments, the application to the communication system has been described as an example, but the voice information analysis apparatus according to the first and second embodiments uses the analysis result of the analysis unit 600 in addition to this. Therefore, it can be applied to various devices such as a device that performs processing such as voice recognition.

[0042]

As described above, according to the present invention, there is provided a voice analysis device capable of efficiently removing noise whose characteristics dynamically change from a voice signal to be subjected to voice analysis. can do.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a sound / silence determination unit according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of an analysis unit according to the first embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a sound / silence determination unit according to a second embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of an analysis unit according to a second embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a communication system according to an embodiment of the present invention.

[Explanation of symbols]

 100 voice information analysis device 200 voice input device 300 A / D converter 400 buffer memory 401 frame data 500 voice / silence determination unit 511 voice / silence determination unit 600 analysis unit 610 FFT data setting unit 620 FFT unit 630 power spectrum conversion Section 631 power spectrum sequence 640 axis conversion section 650 pitch extraction section 651 pitch information 660 noise removal section 670 inverse FFT section 680 normalization section 681 normalized waveform sequence 682 level information 690 noise processing section 700 silence frame setting section 710 noise table 720 Voice / silent part 800 Vector quantizer 801 Vector code 900 Transmitter 1100 Vector inverse quantizer 1200 Combiner 1300 D / A converter 1400 Audio output device

Claims (6)

[Claims] 1. A voice information analysis method for outputting a result of analysis of voices represented by frame data obtained by collecting voice sampling data for a certain period of time, wherein voices other than noise are included in voices represented by each frame data. If it is determined that the voice other than noise is not included, the feature of the noise included in the voice represented by the frame data is extracted from the frame data and stored. , A step of executing a silence processing for outputting information representing the analysis result of a silent voice prepared in advance as the analysis result of the voice indicated by the frame data, and when it is determined that a voice other than noise is included. , The noise feature stored in the previous silence processing is removed from the voice represented by the frame data, and the voice represented by the frame data from which the noise feature is removed is separated. And voice information analyzing method characterized by a step of performing a sound system processing of outputting the result of analysis. 2. A voice information analysis apparatus for outputting an analysis result obtained by analyzing voices represented by frame data obtained by collecting voice sampling data for a certain period of time, wherein a noise table and a voice represented by each frame data have noise. Means for determining whether or not a voice other than noise is included based on the amplitude of the voice represented by the frame data, and the frame data from the frame data when it is determined that a voice other than noise is not included. Means for extracting the characteristics of noise contained in the voice represented by the above, storing it in the noise table, and outputting information representing the analysis result of the silent voice prepared in advance as the voice analysis result indicated by the frame data. And, when it is determined that a voice other than noise is included, the feature of the noise stored in the noise table represents the voice represented by the frame data. And a unit for analyzing the voice indicated by the frame data from which the noise feature has been removed and outputting the analysis result. 3. Pitch information representing a pitch of a voice represented by each frame data and waveform information representing a voice waveform within the pitch are analyzed by analyzing voices represented by frame data obtained by collecting voice sampling data for a certain period of time. A voice information analyzer for outputting amplitude information representing the amplitude of the voice represented by the frame data, and means for obtaining the power spectrum of the voice represented by each frame data, and frame data from the obtained power spectrum. Means for extracting and outputting the pitch information of the voice indicated by, the noise table, and whether or not the voice represented by each frame data includes voice other than noise is determined by the amplitude of the voice represented by the frame data. Based on the power spectrum of the voice represented by the frame data when it is determined that voice other than noise is not included. Means for storing the table, - the Noizute by extracting a feature of noise included in the speech represented by over data
When it is determined that the sound other than noise is not included, as the analysis result of the sound indicated by the frame data,
When it is determined that the means for outputting the waveform information that represents the silent waveform prepared in advance and the sound other than noise are included,
A means for removing the noise feature stored in the noise table from the power spectrum of the voice represented by the frame data, and the noise feature when the voice other than the noise is determined to be included. And a means for extracting the waveform information and the amplitude information from the removed power spectrum by using the pitch information obtained for the corresponding frame data and outputting the extracted information. . 4. A voice represented by frame data obtained by collecting voice sampling data for a certain period of time is analyzed, and pitch information representing a pitch of a voice represented by each frame data and waveform information representing a voice waveform within the pitch. A voice information analyzer for outputting amplitude information representing the amplitude of the voice represented by the frame data, a means for obtaining the power spectrum of the voice represented by each frame data, and frame data from the obtained power spectrum. Means for extracting and outputting the pitch information of the voice indicated by, the noise table, and whether or not the voice represented by each frame data contains voice other than noise. Or a means for making a determination based on the pitch information obtained for the frame data and the amplitude of the voice represented by the frame data. Noise processing for extracting the characteristic of noise contained in the voice represented by the frame data from the obtained power spectrum and storing it in the noise table when it is determined that the voice other than the noise is included. Means and means for outputting waveform information representing a silent waveform prepared in advance as a result of analysis of the voice indicated by the frame data when it is determined that voice other than noise is included, and voice other than noise When it is determined that the noise is included, a means for removing the noise feature stored in the noise table from the power spectrum of the voice represented by the frame data, and a voice other than noise are included. If it is determined that the waveform information and the amplitude information from the power spectrum from which the noise feature is removed,
Means for extracting and outputting the pitch data obtained for the music data using the pitch information. 5. A voice compression code having the voice information analysis apparatus according to claim 3 and 4, and means for quantizing waveform information output from the voice information analysis apparatus and outputting quantized data. Device. 6. A means for inputting voice, a means for sampling the input voice and outputting sampled data, a buffer means for accumulating the sampled data and outputting the frame data in which the sampled data is accumulated for a predetermined time. And a means for transmitting the pitch information, the waveform information, and the amplitude information output from the speech compression encoding apparatus through a wired or wireless transmission path. A communication terminal device characterized by the above.