JPH03254223A - Analog data transmission system - Google Patents

Abstract

PURPOSE:To simplify the constitution of a transmission line and a transmission equipment and to allow a reception side to receive a data with high fidelity by compressing an analog data with a band limit at a transmission side or a transmission line in the case of transmitting the analog data. CONSTITUTION:A transmission analog data 100 is made to a narrow band by a band limiter 10 composed of a low pass filter at the transmission side. A waveform 110 with a high band component eliminated therefrom is sent from a transmitter 12 to a transmission line 14 as a signal whose information quantity is considerably missing while it fundamental component only remains and received by a receiver 16. Then the reception signal is fed to a synthesizer 20 as it is and decoded by a neural network 20 studied for each data processed in advance to decode the lost high frequency component. Thus, the synthesizer 20 synthesizes the received data 200 with a predicted value obtained from the neural network 26 and outputs the result as a synthesis data 210. Thus, while the load of the transmission line or the transmission equipment is relieved, the data transmission with fidelity is attained.

Description

[Detailed Description of the Invention] [Industrial Application Fields] The present invention relates to an analog data transmission system, particularly an improved analog data transmission system that compresses data by band-limiting the transmission signal and reduces the burden on the transmission path. Regarding the method.

[Prior Art] In recent years, the need to transmit audio or image data has increased significantly, and a wide variety of transmission means have been put into practical use, such as communication systems using telephone lines, television communication systems, or various databases. ing.

When handling analog data as such transmission information,
The problem is the bandwidth.

Considering the requirements of the transmission path, it is desirable for analog data to have as narrow a band as possible, and in reality, even though analog data has a wide band, some particularly high band components are discarded due to the band limitation of the transmission path. It has been known that it can get lost.

In addition, various types of media such as magnetic tape are used to store analog data, and analog data is always subject to bandwidth limitations due to the need to reduce the size and weight of such storage or transmission media. Become.

On the other hand, such band limitation causes extremely serious characteristic deterioration for analog data itself, and in normal cases, band limitation results in data loss for high frequency components. It is known that analog data with such high frequencies cut off causes a slowing of the signal waveform, which in turn leads to significant deterioration in sound quality, such as muffled sound or loss of sense of speed at rise.

For example, taking general audio as an example, these analog audio data usually have a high band of up to 3.4 kHz, but using a normal transmission path, the band is limited to, for example, 0.8 kHz, and as a result, The thin high-frequency components that normally precede the sound are lost, resulting in a distorted waveform that becomes significantly dull. Due to such waveform deterioration, although basic frequency information representing the pitch of the sound source remains, the tone becomes muddy and unpleasantly muffled. Furthermore, as mentioned above, the rise of the sound becomes slow, which impairs the sense of speed.

Furthermore, not only does recognition of consonants become partially difficult;
Sometimes the second formant is missing even in the vowel part,
It also caused misunderstandings.

Therefore, in analog data processing, the requirements of the transmission line or transmission equipment and the requirements of data quality always conflict, and it has been extremely difficult to adjust them.

Conventionally, an analog device has been used to restore the band-limited analog data.

In other words, in order to improve the clarity of band-limited audio analog data, for example, Tokagi increases the degree of amplification in the limited frequency band and restores the analog waveform, thereby enhancing data reception through the emphasized amplification effect of Tokagi. On the other hand, restored data close to the original data can be obtained.

The conventional restoration technique using an equalizer has been put into practical use as a signal regeneration method using a linear method, for example, and all of them utilize a small amount of residual data of high-frequency components that are subject to band limitations. It is something to do.

[Problems to be Solved by the Invention] However, the conventional restoration techniques described above have a problem in that received data of sufficient quality cannot necessarily be obtained.

In other words, when using an isoflower, the waveform should be able to be restored using the inverse characteristics of the isoflower determined according to the transmission path characteristics, but in this case, even a small amount of data is transmitted in the desired high band. It is necessary to be present. In other words, if high band component data is completely lost,
Conventional flower vases are almost unable to perform their functions, making data recovery impossible.

Furthermore, even in conventional signal reproduction methods that use linear methods, for example, in the field of audio processing, a perfect pole-zero linear model has not yet been discovered, and originally only audio with a small amount of data can be obtained on the receiving side. In situations where there is no such thing, data recovery is virtually impossible.

In the past, technology has been proposed that uses spectrum estimation technology to restore lost high-frequency component data even when data on the receiving side is completely lost, limited to audio data. Even with estimation, it is difficult to estimate the consonant information of the voice, and furthermore, the phase information possessed by the analog data is completely ignored, resulting in a change in sound quality.

The present invention was made in view of the above-mentioned conventional problems, and its purpose is to provide an improved analog data transmission that enables data recovery with high fidelity on the receiving side even when high-band data is completely lost. The goal is to provide a method.

[Means for Solving the Problem] In order to achieve the above object, the present invention solves the problem that poor analog data retains basic information in a relatively low band,
It is characterized by focusing on the fact that high-band components have fewer inherent differences than low-band components, and as a result, learning can be done relatively easily from low-band components.

For this reason, in the present invention, when transmitting analog data, the transmitted data is actively band-limited, and high-band components are spontaneously dropped.

As a result, the data transmission path or the transmission equipment needs only a narrow band, so it has the advantage of being smaller and lighter, and the transmission density can be significantly improved.

The present invention is characterized in that a restoration process using a neural network is performed on the reception side, thereby adding the missing high-band components at the time of reception.

The neural network is, of course, trained in advance for each analog data to be processed, so that the neural network creates high-band components corresponding to the received low-band analog data and performs high-fidelity restoration. be able to.

That is, according to the present invention, as described above, the high band components in analog data have relatively small inherent differences and can be estimated relatively easily from the low band components, and the high band components are completely Even in the missing state, neural network learning enables high-fidelity restoration.

Of course, the above-mentioned removal of high band components, which is a feature of the present invention, may be achieved by using a band limiter or by using the narrow band characteristics of the transmission path itself.

[Operation] Therefore, according to the present invention, analog data such as audio or images is actively band-limited beforehand on the transmitting side or on the transmission path before being sent.

The neural network provided on the receiving side is trained in advance using pack propagation or the like on the analog data to be processed, and the input/output characteristics of the neural network are determined.

When a neural network trained in this way is supplied with band-limited analog data, the neural network generates high-band data from low-band information, which makes it possible to restore received data that is close to the original data. becomes.

Data restoration using the neural network may be performed by converting the received data itself into restored data, or by restoring only the lost high frequency component from the time difference and combining the received data with this.

[Example] Preferred embodiments of the present invention will be described below based on the drawings.

FIG. 1 shows the overall configuration of a transmission system to which the present invention is applied.

For example, the analog data 100 to be transmitted is shown in the figure as a voice "female's 'a'" having a band of 3.4 kHz, and this analog data 100 to be transmitted is reduced to a band of 0.5 kHz by a band limiter 10 consisting of a low-pass filter on the transmitting side. The band is narrowed to 8 kf (z). Of course, this band-limiting frequency can be arbitrarily selected depending on the transmission line or transmission equipment, and as will be detailed later, experiments have shown that even if the band frequency changes, the present invention No significant effect was seen on the restoring effect.

As shown in the figure, the waveform 110 from which the high band components have been removed is shown as a signal with a large amount of information missing, leaving only the basic pitch.

Such a narrowband signal 110 is transmitted to the transmitter 1 in the embodiment.
2 to the transmission line 14, and on the receiving side, the receiver 1
Received at 6.

In the present invention, such a transmission system can be arbitrarily used regardless of whether it is analog or digital processing, wired or wireless, and the feature of the present invention is that narrow band limitation is applied at the analog signal stage on the transmitting side or transmitter. The objective is to enable data compression of the analog data itself by applying .

Consider that the output of the receiver 16 is approximately the same as the narrowband signal 110 in FIG.

In this embodiment, this received signal is supplied as it is to the synthesizer 20, and is restored by a neural network in order to restore the lost high-frequency components.

The received data 200 is first supplied to the sample circuit 22,
Data is sampled by the sampling clock. In the example, the sampling clock is 1
It has a frequency of 0 kHz, and 40 pieces form one frame. That is, these 40 neural networks have a 4-layer perceptron configuration, and each layer has 40 perceptrons.
The 40 sampling data are allocated to the neural units of each input layer, respectively.

The data sampled as described above is then supplied to the differentiator 24, and in the embodiment, a time difference value is output.

Of course, it is also possible to supply the sampled waveform data itself to the neural network in the present invention, but in the embodiment, the time difference value is used as manual data for the neural network as described above in order to emphasize high bands. It will be done.

Then, the difference data is stored in the neural network 26.
undergoes a predetermined conversion process.

That is, the 22-ral network 26 predicts the error between the received data and the transmitted data according to the conversion rule generated by learning itself, and outputs the value.

In the illustrative embodiment, this predicted value is provided to the other input of the synthesizer 20 described above, so that the synthesizer 20 combines the received data 200 with the predicted value obtained from the neural network 26 and generates the predicted value at 210 in the figure. Output as the composite data shown.

As described above, according to the present invention, data to be transmitted is compressed in advance in an analog state, and this is restored on the receiving side using a neural network, thereby burdening the transmission line or transmission equipment. This has the advantage of enabling high-fidelity data transmission while reducing data transmission.

As is well known, the neural network 26 described above has a two-step process of learning and execution, and according to the present invention, it learns the conversion rules necessary for restoration based on the analog data to be processed, and then performs the reception as described above. As an execution step, data is input and a predicted value according to the learned conversion rule is output.

That is, in the learning stage, the neural network is made to empirically learn a conversion rule for estimating the error between the original transmission data before the band limitation and the received data from the received data degraded due to the band limitation.

Then, in the execution stage, the estimated error with respect to the original transmitted data is output for the unknown received data input, and according to the embodiment, the synthesized waveform is created by adding this predicted error value to the received data as described above. You will get it.

The contents of the neural network 26 will be explained in detail below.

FIG. 2 shows a general neuron model, and shows an engineering model in which the operation of a neuron, usually called a unit, is modeled as an information processing functional element.

The unit processes the signals from the input side using weighted majority logic and produces one output signal.

Then, this output signal is used as an input signal for a plurality of units in the next layer to perform learning and execution processing using a multilayer hierarchy. This model captures two processes as the basic properties of a neuron: a product-sum operation of an input signal and a weight, and a threshold value process for the result of the product-sum operation. Using the sigmoid function, the human output characteristics of the unit can be formulated as follows.

o, -f-(net-)
(2)1) J J pJf
, (x)=1/ (1+exp (-x+θ)
) (3) A neural network is constructed by interconnecting the aforementioned large number of units (neurons), and various variations are created depending on the form of the connection.

FIG. 3 shows the basic structure of the neural network according to the present invention, which has a network structure consisting of four layers: an input layer, two intermediate layers, and an output layer.

Each unit in the hierarchical network is connected in the direction from the human-powered layer to the intermediate layer and from the intermediate layer to the output layer, and there are no connections within each layer or from the output layer to the human-powered layer.

When a human cover turn is given to the human power layer, each unit of the input layer passes the human power signal to the middle layer, and each unit of the middle layer operates based on the above-mentioned equations (1) and (2),
An output signal is supplied to each unit of the output layer.

Finally, the output pattern of the network is output from each unit of the output layer.

Therefore, by always processing the input forward, this network applies a certain transformation rule (data processing method) determined from the weight W of connections between units to the input cover pattern, and outputs that pattern to the output layer. do.

When operating the general neural network described above, the transformation rules described above must be provided, and the network needs to be trained for this purpose. This network learning is the process of determining connections between units such that, for each given input pattern, the output pattern actually output by the network and the desired output pattern (teacher signal) match or are sufficiently similar. The purpose is to find a set of weights W.

As a learning algorithm for such a hierarchical network, the backpropagation method proposed by Rumelhart et al. is known. In the backpropagation method, the error between the actual output pattern and the desired output pattern (teacher signal) is fed back to the hierarchical network, and based on this, the connection weight W and bias value θ are determined as shown in (4) below.
Adaptively auto-adjust based on the formula.

Δ wo, ηδ , i ,
(4)1) Jl pJ p1 δ for the output layer unit is l δ , −(t , −o ,) f ,
(nepj pJ I)j
δpt for the unit of J middle layer is δ−=f−, (net,)ΣδpkWkjpJ
It is calculated as J pj.

As a result of the above, the network learns the association between input and output patterns, and can realize adaptive data processing.

In the present invention, as is clear from FIG. 1 mentioned above,
The neural network is given the time difference value of the received data subject to the band limit, and the difference value is given to each unit of the human layer of the neural network 26 for every frame of a certain length (40 samples in the embodiment). , learning of the network connection weights W proceeds in a direction that re-digests the numerical gap between the signal output from the output layer and the difference between this signal and the band-limited received data.

Therefore, after the above learning, the neural network 26 will have the optimum combination weight W for restoring the band-limited data from which high bands have been removed.

t, ) (5) pJ (6) Figure 4 shows the neural network 2 in this embodiment.
6, the learning and restoring operations are performed on band-limited received data under the following conditions.

Sampling frequency: 10kHz frame length
:4m5 (40 pieces) Frame shift:
2m5 (20 pieces) Number of sample points 26000 pieces (vowels): 2000 pieces (consonants and 10 vowels) Number of accumulated frames: 50 pieces (vowels) = 25 pieces (consonants and 10 vowels) Learning efficiency constant (η) +0.002 (vowels) ):0.
004 (consonant ten vowels) Stabilization coefficient (α): o, c+ configuration, 4-layer berseptron.

(40, 40, 40, 40) Input signal: 013kl (difference between z-band signal Desired output signal: 0.1ikHz band signal and 3.4kHz band difference) Here, the number of sample points is the difference between the z-band signal and the 3.4kHz band signal. Represents the number of points earned.

Here, the first 0.6 seconds are used for learning in the case of only vowels, and the first 0.2 seconds in the case of syllables containing consonants. In addition, the number of accumulated frames refers to the number of frames since the unit error is accumulated for a certain number of frames during learning, and then the error is propagated and the network weights are updated. . By accumulating unit errors, the network can broadly generalize across all data to find the optimal solution, rather than fixating on individual frame data. If this value is too small, the network becomes dependent on individual data, and it takes a lot of time to search for the optimal solution for all data. Conversely, if this value is large, it becomes difficult to capture the fine features of the waveform.

Tables 3 and 4 show the subjective evaluation results (MOS: MEANOPI
NION 5CORE).

Table 3 Subjective evaluation results for sentence utterances (learning speakers) Table 4 Subjective evaluation results for sentence utterances (untrained speakers) In this evaluation, the 3.4kHz band signal is considered to be the original sound (equivalent to 4 points), and 0 points ( Ten people were asked to rate the sound quality on a five-point scale from 4 points (very poor sound quality) to 4 points (same as the original sound), and the average score was given. Comparing the signals in each band and the synthesized sound, the signal synthesized using the 0.8 kHz band signal exceeds the received signal in the 0 and 8 kHz bands for both trained and untrained speakers. Of course, 2,0k
A better score was obtained than the signal in the Hz band. This demonstrated the effectiveness of predicting loss information and expanding the signal band.

In this example, it is assumed that a signal in the 0 and 11 kHz band is input, but in order to consider fluctuations in the frequency characteristics of the transmission path, a signal synthesized based on a signal whose band is limited to 0.7 kHz in the learning speaker is used. Signals were also evaluated. The results are also listed at the bottom of Table 3. Although there was some deterioration compared to the sound synthesized based on the 0.8 kHz band signal mentioned above, a score close to that of the 2.0 kHz band signal was obtained. Therefore, it has been found that the method proposed by the present invention can be applied even to slight variations in the frequency characteristics of the transmission path.

[Effects of the Invention] As explained above, according to the present invention, when transmitting analog data, the analog data is compressed by band limiting on the transmitting side or on the transmission path, thereby reducing the speed of the transmission path and transmission equipment. It is possible to make the configuration extremely simple.

Furthermore, on the receiving side, data is restored in accordance with preset conversion rules using a neural network, thereby making it possible to receive data with extremely high fidelity.

[Brief explanation of drawings]

Figure 1fi1 is a schematic configuration diagram showing a preferred embodiment of the analog data transmission system according to the present invention, Figure 2 is an explanatory diagram showing the principle of a neuron used in the present invention, and Figure 3 is a neural network used in the present invention. FIG. 4 is a diagram illustrating the operation of the neural network according to the present invention. 10...Band limiter 14 Transmission line 6 Neural network

Claims (1)

[Claims] An analog data transmission method characterized by transmitting analog data after removing high-band components, and restoring the high-band components on the receiving side using a neural network trained in advance according to data characteristics.