JPH02289899A - Voice detection system - Google Patents

Abstract

PURPOSE:To detect the presence of a voice at a high detection rate even when the amplitude of a noise is large by analyzing the frequency of an input signal by a preprocessing part, inputting the analysis result to a neural network, and deciding whether or not the input signal contains a voiced sound from the output of the neural network. CONSTITUTION:The preprocessing part 10 analyzes the frequency of the input signal which is inputted from a signal input part, the analysis result of the preprocessing part 10 is inputted to the nueral network 20, and a decision circuit 30 decides whether or not the input signal contains the voiced sound from the output of the neural network 20. Consequently, even when the amplitude of the noise is large and an influence upon the detection of a voice is large, the presence of the voice in noisy environment can be detected at a high detection rate and short-time processing is easily performed.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a voice detection method.

[Prior Art] Conventionally, there are many methods for detecting the presence of speech in a noisy environment. Although it is used for recognition preprocessing, it is difficult to deploy it for general purposes under high noise conditions, such as initiating a voice response from a hands-free telephone when the incoming bell is ringing. could not. Incidentally, as a method for simply detecting the presence of voice in a noisy environment, there is a method of detecting the number of times an input signal crosses a reference axis within a fixed time interval.

However, the method using the conventional voice detection method described above uses the assumption that the amplitude of the noise is small compared to the amplitude of the voice. In this case, the presence of voice cannot be detected. Therefore, the present applicant has developed a voice detection method that can easily detect the presence of voice in a noisy environment, based on the number of reference axis crossings of the input signal and the peak value (the amount of non-dimensionalization of the amplitude level of the waveform). We are proposing a method for determining whether an input signal contains voiced sounds by pattern recognition, by calculating them as feature quantities, and comparing the calculation results with predefined dictionary data for voiced sounds and specific noises. [Problems to be solved by the invention] Silently, the above conventional voice detection method has the following problems:
■There is a problem. ■It is impossible to recognize patterns assuming all noises other than the specific noise (for example, a bell sound) used when creating dictionary data, and the detection rate is low in environments with unexpected noise. ■In conventional voice detection methods, it is necessary to use complex features as described above in order to ensure a detection rate above a certain level, but this requires a complex processing device and requires a relatively long processing time. The present invention provides a voice detection method that can detect the presence of voice in a noisy environment with a high detection rate even when the amplitude of noise is large and has a large influence on voice detection, and that can be easily processed in a short time. With the goal. [Means for solving the problem] The present invention according to claim 1 analyzes the frequency of an input signal input from a signal input section in a preprocessing section, and inputs the analysis result by the preprocessing section to a neural network. , the output of the neural network is used to determine whether the input signal contains voiced sounds. According to the second aspect of the present invention, the analysis result by the preprocessing section is the frequency characteristic of the input signal. The present invention according to claim 3 is characterized in that the analysis result by the preprocessing section is a temporal change in the frequency characteristics of the input signal within a certain period of time. The analysis result by the section is the average frequency characteristic of the input signal. According to a fifth aspect of the present invention, the analysis result by the preprocessing section is a temporal change in the average frequency characteristic of the input signal within a certain period of time. According to a sixth aspect of the present invention, the analysis result by the preprocessing section is an average frequency characteristic of an input signal and an average pitch frequency. The present invention according to claim 7 is such that the analysis result by the preprocessing section is a temporal change in the average frequency characteristic of the input signal within a certain period of time, and a temporal change in the average pitch frequency. It is.

According to an eighth aspect of the present invention, the analysis result by the preprocessing section is an average frequency characteristic of an input signal with high frequency emphasis. According to a ninth aspect of the present invention, the analysis result by the preprocessing section is a temporal change in the average frequency characteristic within a certain period of time of an input signal with high frequency emphasis. According to a tenth aspect of the present invention, the neural network is a hierarchical neural network. [Function] According to the present invention described in each of claims 1 to 9, there are the following effects (1) to (4). In the present invention, voiced sounds (sounds that involve vibration of the vocal cords, such as vowels, semi-vowels, and nasal sounds, and almost all sounds produced by humans include voiced sounds) are defined as speech. ．． ■Neural networks can be additionally trained as needed during the system operation stage after the network has been constructed through learning, which will be described later. Therefore, even for voice detection in noisy environments that was not anticipated during the learning stage for network construction, a high detection rate can be achieved by additionally learning this at any time during the operation stage. ■Since the "frequency analysis results of the input signal" are used as the input to the neural network, the preprocessing to obtain the input is simpler than the conventional complex feature extraction, and the time required for this preprocessing is is short enough. ■Neural networks are, in principle, capable of simple and quick arithmetic processing for the entire network. ■In principle, each unit that makes up a neural network operates independently, and parallel arithmetic processing is possible. Therefore, calculation processing is quick. ■With the above ■~■, voice detection processing can be easily performed in a short time without using a complicated processing device. Further, according to the present invention as set forth in claim 10, in addition to the effects (1) to (2) above, there is the following effect (2). ■For hierarchical neural networks, currently,
A simple learning algorithm (pack-broadcasting) has been established as described below, and it is easy to create a neural network that can achieve a high detection rate. [Example] FIG. 1 is a schematic diagram showing an example of a voice detection system to which the present invention is applied, FIG. 2 is a schematic diagram showing a neural network, FIG. 3 is a schematic diagram showing a hierarchical neural network, and FIG. Figure 4 is a schematic diagram showing the structure of the unit. Before explaining specific embodiments of the present invention, the configuration of the neural network and the learning algorithm will be explained. (1) Due to its structure, neural networks
It can be roughly divided into two types: the hierarchical network shown in Figure (A) and the interconnected network shown in Figure 2 (B). Although the present invention may be configured using either of these networks, the hierarchical network is more useful because a simple learning algorithm has been established as described below. (2) Network structure A hierarchical network has a hierarchical structure consisting of an input layer, a middle layer, and an output layer, as shown in Figure 3. Each layer consists of one or more units. The only connections are forward connections such as input layer → middle layer → output layer, and there are no connections within each layer. (3) Unit structure The unit is a model of a neuron in the brain and has a simple structure, as shown in Figure 4. It receives input from other units, sums it up, transforms it using a certain rule (conversion function), and outputs the result. Each connection with another unit is given a variable weight that represents the strength of the connection. (4) Learning (pack-broadcasting) Network learning is to bring the actual output closer to the target value (desired output), and in terms of the number of transformations and weights of each unit shown in Figure 4. Learning is performed by changing . Specifically, the target values are set to ``1'' for voiced sounds and ``0'' for noise, and according to the following ■~■. ■Apply the preprocessing of the present invention only to voiced sounds, and input the preprocessing results to the neural network. Then, modify the conversion function and weight of each unit so that the output of the neural network approaches the target value. ■Apply the preprocessing of the present invention only to the noise, and input the preprocessing results to the neural network. Then, modify the conversion function and weight of each unit so that the output of the neural network approaches the target value. ■It is also possible to perform learning with input signals that include voiced sounds and noise. The target value in this case is "1" for voiced sound.

Further, as a learning algorithm, for example, Rumel
hart, D. E. ．． McCelland, J.
L. and thePDP Research Group
oup, PARALLEL DISTRIBIJTE
DPROCESSING, Seven he MIT Press
, 1988. You can use pack propagation as described in . (5) Evaluation After constructing a network that can ensure a constant detection rate through the above learning, input the preprocessed unknown input signal into the neural network. If the output result of the neural network is close to rlJ, it is a voiced sound, "0".
If it is close to , it is judged as noise. Hereinafter, specific examples of the present invention will be described. still,
The detection system 1 of this embodiment is composed of a combination of an n-channel bandpass filter 10, a neural network 20, and a determination circuit 30 (see FIG. 1). TA) The input signals in the learning stage for network construction are, for example, ■ the steady part of the voiced sound "a" (the part excluding the rising and falling parts of the signal), and ■ the bell sound (specific noise). Suppose that Note that the specific noise employed in this learning stage is not limited to the bell sound, but may be any noise that is expected to occur in the environment in which the system will be used. (B) Pre-processing As shown in FIG. 1, the input signal waveform is passed through a bandpass filter 10 with multiple (n) channels to obtain the frequency characteristics of the input signal.

(C) Processing and judgment by neural network ■Preprocessing results (output of bandpass filter 10)
As shown in FIG. 1, input is made to a three-layer hierarchical neural network 20. The input layer 21 is composed of n units corresponding to n channels of preprocessing. Output layer 22
is composed of one unit, and the target value is ``1'' for voiced sounds and rOJ for noise. (2) The output of the neural network 20 is input to the determination circuit 30, and it is determined according to the output value of the output layer 22 whether the input signal includes a voiced sound. However, in implementing the present invention, the output of the neural network 20 is not mechanically subjected to judgment processing such as in the judgment circuit 30;
The judgment process may be performed using human intellect after obtaining the output. ■Using the pack propagation of the learning algorithm described above, we construct a network that can guarantee a constant detection rate by learning until the error in the output relative to the input converges to a certain level. ■Using the neural network 20 constructed in (■) above, the presence of speech is detected in any noisy environment. At this time, in the actual system operation site, if it is thought that the influence of background noise that was not anticipated during the learning stage for network construction is large, additional learning is performed under the actual usage environment, and as a result, the The network 20 can be improved to better suit the usage environment. (DJ experiment: Using the detection system 1 described above, a voice detection experiment was conducted. As a result, the detection rate was found to be 99%.

Next, the operation of the above embodiment will be explained. According to the above detection system 1, there are the following effects. - As mentioned above, the neural network 20 can perform additional learning as needed during the system operation stage after the network has been constructed through initial learning. Therefore, even for voice detection in noisy environments that was not anticipated during the learning stage for network construction, a high detection rate can be achieved by additionally learning this at any time during the operation stage. ■Since the "frequency characteristics of the input signal" is used as the input to the neural network 20, the preprocessing to obtain the input is simpler than the conventional complex feature extraction, and the time required for this preprocessing is is short enough. - In principle, the neural network 20 has simple and quick calculation processing for the entire network. ■The neural network 20 has, in principle, each unit that constitutes it operating independently, and is capable of parallel arithmetic processing. Therefore, calculation processing is quick.

■Doki ■ to ■ allow voice detection processing to be easily performed in a short time without using a complicated processing device.

■Because we used a hierarchical neural network 20,
By using a simple learning algorithm (pack-broadcasting) that has already been established, it is possible to easily create a neural network that can achieve a high detection rate. Note that in implementing the present invention, the analysis result by the preprocessing section may be a temporal change in the frequency characteristics of the input signal within a certain period of time.

Further, in implementing the present invention, the analysis result by the preprocessing section may be the average frequency characteristic of the input signal. Furthermore, the analysis result by the preprocessing section may be a temporal change in the average frequency characteristic of the input signal within a certain period of time. The detection system equipped with the preprocessing section in this case is as shown in FIG. That is, the detection system 1 is configured by combining an n-channel bandpass filter 10, an averaging circuit 15, a neural network 20, and a determination circuit 30. The preprocessing is performed according to the following ■ and ■. ■The input signal waveform is temporally divided equally into four blocks as shown in Figure 8. ■As shown in FIG. 5, the input signal waveform of each block is passed through a bandpass filter 1o of multiple (n) (in this example, n=8) channels, and each block, that is, each fixed time period, is filtered as shown in FIG. Obtain frequency characteristics as shown in each of A) to (D). At this time, the output of the bandpass filter 1o is averaged by the averaging circuit 15 for each block, that is, for each fixed time period. Through the above preprocessing, the temporal change in the average frequency characteristics of the input signal within a certain period of time can be obtained. Further, in implementing the present invention, the analysis results by the preprocessing unit may be the average frequency characteristics and average pitch frequency of the input signal. Furthermore, the analysis result by the preprocessing unit may be a temporal change in the average frequency characteristic of the input signal within a certain period of time, and a temporal change in the average pitch frequency. Note that the pitch frequency of speech is the reciprocal of the repetition period (pitch period) of vocal cord waves. And as input to the neural network,
When pitch frequency, which is a basic parameter of the vocal cords that varies among individuals, is added, the detection rate of voiced sounds can be improved for unspecified speakers. The detection system equipped with the preprocessing section in this case is as shown in FIG. That is, the detection system 1 is configured by combining an n-channel bandpass filter 10, a pitch extractor 11, an averaging circuit 15, a neural network 20, and a determination circuit 30. The preprocessing is performed according to the following ■ and ■. ■The input signal is temporally equally divided into four blocks as shown in Figure 8. ■Input signal waveforms are input to multiple (n) (as shown in Figure 6).
In this embodiment, each block, that is, each fixed time period, is passed through a bandpass filter 10 of n=8 channels, as shown in FIG.
Obtain frequency characteristics as shown in each of A) to (D). Further, in parallel with the processing by the bandpass filter 10, the input signal waveform is passed through a pitch extraction section 11 to obtain a pitch frequency for each block, that is, for each fixed time period. At this time, the respective outputs of the bandpass filter 10 and the pitch extractor 11 are averaged by the averaging circuit 15 for each block. Through the above preprocessing, a temporal change in the average frequency characteristic and a temporal change in the average pitch frequency of the input signal within a certain period of time can be obtained.

Further, in implementing the present invention, the analysis result by the preprocessing unit may be the average frequency characteristic of the input signal with high frequency emphasis. Furthermore, the analysis result by the preprocessing unit may be a temporal change in the average frequency characteristic within a certain period of time of the high-frequency emphasized input signal. Note that high frequency enhancement is to compensate for the average slope of the spectrum of the audio waveform and prevent energy from concentrating in the low frequency range. In this case, the detection system equipped with a pre-processing section is
It will look like Figure 7. That is, the detection system 1 includes a high frequency emphasizing section 10A, an n-channel bandpass filter 10,
It is composed of a combination of an averaging circuit 15, a neural network 20, and a determination circuit 30. And the pre-processing is
According to ■ and ■ below. ■The input signal is temporally equally divided into four blocks as shown in FIG. 8. ■The input signal waveform is passed through the high-frequency emphasizing section 10A consisting of a high-frequency emphasizing filter as shown in FIG. Emphasizes high frequencies. Next, the audio waveform after high-frequency emphasis is passed through a bandpass filter 10 with multiple (n) (n=8 in this example) channels, and each block, that is, each fixed time period, is filtered as shown in FIG.
) to (D) are obtained. At this time, the output of the bandpass filter 10 is averaged by an averaging circuit 15 for each block. Through the above preprocessing, it is possible to obtain the temporal changes in the average frequency characteristics of the high-frequency emphasized input signal within a certain period of time. Note that the high frequency enhancement operation of the present invention is performed not before the input to the band pass filter 10 as described above, but after the input to the band pass filter 10.
It may be applied after the output from . [Effects of the Invention] As described above, according to the present invention, even when the amplitude of noise is large and the influence on speech detection is large, the presence of speech in a noisy environment can be detected with a high detection rate and easily. It is possible to obtain a voice detection method that can process questions in a short time.

[Brief explanation of drawings]

Fig. 1 is a schematic diagram showing an example of a voice detection system to which the present invention is applied, Fig. 2 is a schematic diagram showing a neural network, Fig. 3 is a schematic diagram showing a hierarchical neural network, and Fig. 4 is a schematic diagram showing a unit. FIG. 5 is a schematic diagram showing another example of the voice detection system. FIG. 6 is a schematic diagram showing another example of the voice detection system. FIG. 7 is a schematic diagram showing another example of the voice detection system. FIG. 8 is a schematic diagram showing an input signal, and FIG. 9 is a schematic diagram showing an output of a bandpass filter. DESCRIPTION OF SYMBOLS 1... Detection system, 10... Bandpass filter, 10A... High frequency emphasis section, 11... Pitch extraction section, 15... Averaging circuit, 20... Neural network, 21... - Input layer, 22... Output layer, 30... Judgment circuit. Patent applicant Sekisui Chemical Co., Ltd. Representative Kaoru Hirota (A) Figure 2 Figure 3 Human power pattern (B) Figure 4

Claims (10)

[Claims] (1) The input signal input from the signal input section is frequency-analyzed by the pre-processing section, the analysis result by the pre-processing section is input to the neural network, and the output of the neural network is used to determine whether the input signal contains voiced sounds. Voice detection method to judge. (2) The audio detection method according to claim 1, wherein the analysis result by the preprocessing section is a frequency characteristic of the input signal. (3) The audio detection method according to claim 2, wherein the analysis result by the preprocessing section is a temporal change in frequency characteristics of the input signal within a certain period of time. (4) The voice detection method according to claim 1, wherein the analysis result by the preprocessing section is an average frequency characteristic of the input signal. (5) The audio detection method according to claim 4, wherein the analysis result by the preprocessing section is a temporal change in an average frequency characteristic within a certain period of time of the input signal. (6) The audio detection method according to claim 1, wherein the analysis result by the preprocessing section is an average frequency characteristic and an average pitch frequency of the input signal. (7) The audio detection method according to claim 6, wherein the analysis result by the preprocessing section is a temporal change in an average frequency characteristic of the input signal within a certain period of time, and a temporal change in an average pitch frequency. (8) The audio detection method according to claim 1, wherein the analysis result by the preprocessing section is an average frequency characteristic of an input signal with high frequency emphasis. (9) The audio detection method according to claim 8, wherein the analysis result by the preprocessing section is a temporal change in an average frequency characteristic within a certain period of time of the high-frequency emphasized input signal. (10) The voice detection method according to any one of claims 1 to 9, wherein the neural network is a hierarchical neural network.