JPH02273799A - Speaker recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speaker recognition method suitable for identifying or verifying a speaker from input speech using an online terminal such as an electric lock or an IC card.

[Prior Art] Speaker recognition refers to recognizing a speaker from input speech, and there are two forms: speaker identification and speaker verification.

Speaker identification is to determine which of the registered speakers the input voice belongs to, and speaker verification is to determine whether the input voice can be considered to be the voice of the registered speaker. That's true.

Conventional speaker recognition methods include, for example, the Japanese Patent Publication No. 56-13956.
Publication, Journal of the Institute of Electronics and Communication Engineers 73/11 Vol. 56
-A No. 11 (Reference 1), according to the following procedure.

■Extract features related to the speaker included in the input voice.

(2) Calculate the distance between the standard pattern extracted in advance in the same manner as (2) above and the feature amount extracted in (2) above.

■In the case of speaker identification, the speaker of the standard pattern with the minimum distance above is determined to be the speaker of the input voice, and in the case of speaker verification, a threshold value is set in advance for the above distance, and the distance is set to that value. If it is smaller than , it is determined that the input voice is from the speaker of that standard pattern.

At this time, in Japanese Patent Publication No. 13956/1983, the correlation between spectral parameters calculated from the input voice, the average value and standard deviation of each parameter are used as the feature amounts.

[Problems to be Solved by the Invention] However, the conventional speaker recognition method described above has the following problems.

■As time passes from the time the standard pattern was created, the recognition rate deteriorates. For example, in the experimental results described in Reference 1, the speaker identification rate deteriorated from 100% to 85% after three months had passed since the standard pattern was created, and the speaker matching rate decreased to 99%.
% to 91%.

■Real-time processing is difficult. In other words, in conventional speaker recognition methods, it is necessary to use complex features to ensure a recognition rate above a certain level, but extracting complex features requires a complex processing device. It also takes a lot of time.

SUMMARY OF THE INVENTION An object of the present invention is to obtain a speaker recognition method that exhibits extremely little deterioration in recognition rate over time and that can be easily processed in real time.

[Means for Solving the Problems] The present invention as set forth in claim 1 is a speaker recognition method for recognizing a speaker from input speech using a neural network, and includes high-frequency emphasis as an input to the neural network. The average frequency characteristics of the processed audio waveforms are used.

According to a second aspect of the present invention, a temporal change in an average frequency characteristic within a certain period of time of a high-frequency emphasized audio waveform is used as an input to the neural network.

According to a third aspect of the present invention, the neural network is a hierarchical neural network.

Note that high-frequency emphasis means compensating for the average slope of the spectrum of the audio waveform to prevent energy from concentrating in the low-frequency range.

[Function] According to the present invention as set forth in claim 1, there are the following effects (1) to (4).

■Deterioration of recognition rate over time is extremely small. This means that
This is confirmed by the experimental results described below,
It is presumed that this is because a neural network can have a structure that is less susceptible to fluctuations due to differences in audio timing.

■As the input to the neural network is the "average frequency characteristic of the high-frequency emphasized audio waveform," the preprocessing required to obtain the input is faster than conventional complex feature extraction. It is simple, and the time required for this preprocessing is short.

That is, the average slope of the spectrum of the speech waveform is common to all speakers and is unrelated to the speaker's recognition. However, when inputting a speech waveform that has not been compensated for the average slope of the spectrum to a neural network as is, the neural network extracts the feature of the average slope of the spectrum during learning, and the speech becomes distorted. It takes time to extract the peaks and valleys of the spectrum necessary for human recognition. On the other hand, when high-frequency emphasis is applied to the input to a neural network, it is possible to compensate for the average slope of the spectrum that is common to all speakers and is unrelated to recognition, but that affects learning, which speeds up the learning process. becomes faster.

(2) In principle, neural networks allow simple and quick calculation processing of 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.

(2) With the above (2) to (4), speaker recognition processing can be easily performed in real time without using a complicated processing device.

Further, according to the present invention as set forth in claim 2, in addition to the effects (1) to (2) above, there is the following effect (2).

■As input to the neural network, we use the "temporal change in the average frequency characteristics within a certain period of time of the high-frequency emphasized audio waveform", so the processing in the neural network is simple and the time required for this processing is It's shorter than that and it's enough.

Further, according to the present invention as set forth in claim 3, 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 propagation) as described below has been established, and a neural network that can achieve a high recognition rate can be easily formed.

[Example] Fig. 1 is a schematic diagram showing an example of a speaker recognition system to which the present invention is applied, Fig. 2 is a schematic diagram showing input speech, and Fig. 3 is a schematic diagram showing the output of a bandpass filter. FIG. 4 is a schematic diagram showing a neural network, FIG. 5 is a schematic diagram showing a hierarchical neural network, and FIG. 6 is a schematic diagram showing a unit structure.

Prior to describing specific embodiments of the present invention, the configuration of the neural network and the learning algorithm will be described.

(1) Due to its structure, neural networks are
It can be roughly divided into two types: a hierarchical network shown in FIG. 4(A) and an interconnected network shown in FIG. 4(B). The present invention may be constructed using either of these networks, but 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, an intermediate layer, and an output layer, as shown in FIG.

Each layer is composed 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) Structure of the unit The unit is a model of a neuron in the brain and has a simple structure as shown in FIG. 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 propagation) Network learning is to bring the actual output closer to the target value (desired output). Learning is performed by changing the

Further, as a learning algorithm, for example, Rumel
hart, D.E., McClelland, J.
L, and thePDP Re5earch Gr.
oup, PARALLEL DISTRIBUTED
PROCESSING, the MIT Press
, 1986. Pack propagation described in (Reference 2) can be used.

Hereinafter, specific examples of the present invention will be described. still,
The recognition system 1 of this embodiment includes an n-channel high frequency emphasizing section 10A, a bandpass filter 10, and an averaging circuit 15.
, a neural network 20, and a determination circuit 30 (see FIG. 1).

(A) The learning word was "Tadaima", the input word was "Tadaima", and the time difference between the input word and the learning word was O to 3 months.

(B) Speaker ■ Regarding speaker identification, nine registered speakers were identified.

■ Regarding speaker verification, there were 9 registered speakers and 27 learned impostors.
The number of uneducated impostors was 10.

(C) Preprocessing ■ The input voice "Tadaima" is temporally equally divided into four blocks as shown in FIG.

(2) As shown in FIG. 1, the audio waveform is passed through a high-frequency emphasizing section 10A consisting of a high-frequency emphasizing filter to perform high-frequency emphasis.

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 filtered for each block, that is, for each fixed time period, as shown in FIG.
) to (D) are obtained.

At this time, the output of the bandpass filter 0 is averaged by the averaging circuit 15 for each block.

Note that the high frequency enhancement operation of the present invention is performed not before the input to the band pass filter 0 as described above, but after the input to the band pass filter 0.
It may also be applied after the output from.

(D) Processing and judgment by neural network ■Results of preprocessing (bandpass filter 10 for each block
, the output of the averaging circuit 15) are input to a three-layer hierarchical neural network 20 as shown in FIG. The input layer 21 is composed of 4×n units (n=8.32 units in this embodiment) corresponding to 4 blocks of preprocessing and n channels. The output layer 22 is composed of nine units corresponding to each registered speaker for speaker identification, and two units corresponding to the registered speaker group and other groups for speaker verification.

■The output of the neural network 20 is input to the determination circuit 30, and the unit that outputs the maximum value among the output values of each unit of the output layer 22 is identified by the recognition unit 31 and the verification recognition unit 3.
2, and identification or verification is performed as a result. However, in implementing the present invention, the output of the neural network 20 is not mechanically judged by the judgment circuit 30, but is judged by the human intellect that has obtained the output of the neural network 20. Also good.

(2) By using the pack propagation of the learning algorithm described above, the network is trained 30,000 times until the output error with respect to the input converges to a constant level, and a network that can guarantee a constant recognition rate is constructed. It should be noted that learning was performed so that the output would be "1" for the unit corresponding to each input speaker and "0" for the others.

(E) Experiment Using the recognition system 1 described above, speaker identification and speaker matching were experimented.

(a) Speaker identification input speech is ■ “Tadaima” immediately after learning of 9 registered speakers,
■It was named "Tadaima" after 3 months with 9 registered speakers.

As a result, while the identification rate was 100% immediately after learning, the identification rate was 98% after 3 months, and it was observed that the deterioration of the identification rate over time was extremely small.

(b) The input voices for speaker verification are ■ "Tadaima" immediately after learning by 9 registered speakers and 27 learning impostors, ■ "Tadaima" after 3 months by 9 registered speakers and 27 learning impostors, ■Unlearned impostor 1o name “
"Tadaima".

As a result, while the matching rate was 100% immediately after learning, the matching rate was 99% after 3 months, and it was observed that there was very little deterioration in the matching rate over time.

(c) Processing speed For the input speech "Tadaima" once, the time required for identification or verification was less than 1 second, and it was confirmed that processing could be performed extremely quickly.

That is, in the recognition system 1, the above (a),
As shown in the result (b), the recognition rate is high, and the deterioration of the recognition rate over time is extremely small.

Further, in the recognition system 1, as shown in the result (c) above, the speaker recognition process can be performed quickly without using a complicated processing device, and can be easily performed in real time.

In implementing the present invention, the input to the neural network is not the temporal change in the average frequency characteristic within a certain period of time of the high-frequency emphasized audio waveform, but simply "high-frequency emphasized It is also possible to use the average frequency characteristic of the processed audio waveform.

[Effects of the Invention] As described above, according to the present invention, it is possible to obtain a speaker recognition method in which the deterioration of the recognition rate over time is extremely small and which can be easily processed in real time.

[Brief explanation of drawings]

Fig. 1 is a schematic diagram showing an example of a speaker recognition system to which the present invention is applied, Fig. 2 is a schematic diagram showing input speech, Fig. 3 is a schematic diagram showing the output of a bandpass filter, and Fig. 4 is a schematic diagram showing an example of a speaker recognition system to which the present invention is applied. Fig. 5 is a schematic diagram showing a neural network; Fig. 5 is a schematic diagram showing a hierarchical 221 neural network; Fig. 6 is a schematic diagram showing a unit structure. DESCRIPTION OF SYMBOLS 1... Recognition system, 10A... High frequency emphasis part, 10... Band pass filter, 15... Averaging circuit, 20... Neural network, 21... Input layer 22... Output Layer: 30: Judgment circuit; 31: Identification recognition section; 32: Verification recognition section. Patent applicant: Sekisui Chemical Co., Ltd. Representative Hiroshi 1) Kaoru Figure 2 Time construction Figure 3 (A) (B) (C) (D) Figure 4 Figure 5 De-in cover pattern Figure 6 Unisoto

Claims (3)

[Claims] (1) A speaker recognition method that uses a neural network to recognize speakers from input speech, which uses the average frequency characteristics of high-frequency emphasized speech waveforms as input to the neural network. method. (2) The speaker recognition method according to claim 1, wherein a temporal change in an average frequency characteristic within a certain period of time of a high-frequency emphasized speech waveform is used as an input to the neural network. (3) The speaker recognition method according to claim 1 or 2, wherein the neural network is a hierarchical neural network.