WO2020137641A1 - Restoration device, restoration method, and program - Google Patents

Restoration device, restoration method, and program Download PDF

Info

Publication number
WO2020137641A1
WO2020137641A1 PCT/JP2019/049084 JP2019049084W WO2020137641A1 WO 2020137641 A1 WO2020137641 A1 WO 2020137641A1 JP 2019049084 W JP2019049084 W JP 2019049084W WO 2020137641 A1 WO2020137641 A1 WO 2020137641A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
complexity
maximum number
clip signal
program
Prior art date
Application number
PCT/JP2019/049084
Other languages
French (fr)
Japanese (ja)
Inventor
江村　暁
Original Assignee
日本電信電話株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 日本電信電話株式会社 filed Critical 日本電信電話株式会社
Publication of WO2020137641A1 publication Critical patent/WO2020137641A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G10L21/0216—Noise filtering characterised by the method used for estimating noise
    • G10L21/0232—Processing in the frequency domain
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks

Definitions

  • the present invention relates to a technique for restoring a signal before clipping from a signal after clipping.
  • the part where the signal amplitude is larger than the input/output range of the device is clipped to a constant value. Clipping can occur in a wide range of situations, for example, when obtaining a signal from a sensor, when outputting a signal to some device, when inputting an analog signal into an A/D converter for digitization, and the like. Therefore, studies have been made to restore the signal waveform before clipping from the clipped signal.
  • Non-Patent Document 1 a method called SPADE (SParse Audio DEclipper) has been proposed (Non-Patent Document 1).
  • SPADE Pulse Audio DEclipper
  • Each sample of the signal before and after clipping has the relationship of Formula (1).
  • the signal sample after clipping belongs to one of the signal sample S + that is clipped at the upper limit, the signal sample S r that is not clipped, and the signal sample S ⁇ that is clipped at the lower limit.
  • the dictionary matrix D is first defined. Then, paying attention to the signal expression vector z obtained by multiplying the inverse matrix D ⁇ 1 of the dictionary matrix D by the signal vector x, the complexity of the signal is determined by the number of non-zero elements in z, that is, the L 0 norm of z
  • the dictionary matrix D a DFT matrix (Discrete Fourier Transform Matrix), a DCT matrix (Discrete Cosine Transform Matrix), or the like is used.
  • the complexity of a signal before clipping (hereinafter, also referred to as “complexity”) is k, and a predetermined update amount s is an initial value of the complexity k.
  • the input signal that is, the signal y after clipping is converted into a signal expression vector z at D ⁇ 1 . leaving k elements in z from the larger absolute value, the other value by zero, signal representation vector z complexity k - converted to.
  • the estimated signal vector x - is the estimation result of the signal vector x before clipping at this stage.
  • a signal expression vector z ⁇ satisfying the following two conditions is obtained (corresponding to step 3 in Table 1 below).
  • Condition 1 The clipped Dz ⁇ matches y.
  • z ⁇ and z - distance is the minimum.
  • the -SPADE is used in combination with normal frame signal processing. That is, as shown in FIG. 1, the input signal after clipping is divided into frames of a certain length with overlapping in the frame dividing unit 11, and after windowing processing is performed on each frame, the waveform restoring unit 12 SPADE processing is applied. Then, the frame synthesizing unit 13 applies the frame synthesizing process to the processing result, and the restored signal before the clip is obtained.
  • Non-Patent Document 3 the z Step 3 of Table 1 - process of estimating the z ⁇ from the projection Using Will be realized in.
  • the object of the present invention is to realize a technique capable of surely performing waveform restoration even in a situation where the usable calculation amount is limited, in view of the above technical problems.
  • the restoration device repeatedly estimates a pre-clip signal of a predetermined complexity from a clipped signal while increasing the complexity by a predetermined update amount.
  • a restoration device for restoring the pre-clip signal a parameter determination unit that determines the maximum number of repetitions based on the post-clip signal, and an estimation that repeatedly estimates the pre-clip signal with the maximum number of repetitions as the maximum number of estimations And a part.
  • the waveform restoration technology of the present invention it becomes possible to reliably perform the waveform restoration process even in a situation where the usable calculation amount is limited.
  • FIG. 1 is a diagram illustrating a functional configuration of a conventional waveform restoration device.
  • FIG. 2 is a diagram illustrating a functional configuration of the waveform restoration device according to the embodiment.
  • FIG. 3 is a diagram illustrating a processing procedure of the waveform restoration method of the embodiment.
  • the waveform restoration device of the first embodiment first estimates the complexity (sparseness) of a signal before clipping from a signal after clipping using a deep neural network (DNN: Deep Neural Network) (see Reference Document 1). To do.
  • the maximum number of iterations of the estimation process is designated in advance as a limit index of the amount of calculation, and the update amount s of complexity k is controlled based on the estimated degree of sparsity. This makes it possible to reliably perform the waveform restoration process even in a situation where the amount of calculation that can be used is limited.
  • DNN Deep Neural Network
  • the waveform restoration device 10 (hereinafter, also referred to as “restoration device”) of the first embodiment, as illustrated in FIG. 2, a frame division unit 11, a waveform restoration unit 12 (hereinafter, also referred to as “estimation unit”), and In addition to the frame synthesis section 13, a sparse degree estimation section 21 and a parameter determination section 22 are provided.
  • the waveform restoration apparatus 10 implements the waveform restoration method of the first embodiment by performing the processing of each step illustrated in FIG.
  • the waveform restoration device 10 is, for example, a special program configured by reading a special program into a known or dedicated computer having a central processing unit (CPU: Central Processing Unit), a main storage device (RAM: Random Access Memory), and the like. It is a device.
  • the waveform restoration device 10 executes each process under the control of the central processing unit, for example.
  • the data input to the waveform restoration device 10 and the data obtained by each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out to the central processing unit as necessary. It is used for other processing.
  • At least a part of each processing unit of the waveform restoration device 10 may be configured by hardware such as an integrated circuit.
  • step S11 the frame dividing unit 11 divides the input clipped signal into frames.
  • the frame division unit 11 sends the clipped signal frame y to the waveform restoration unit 12 and the sparsity degree estimation unit 21.
  • step S21 the sparseness estimation unit 21 obtains a signal expression vector of the clipped signal frame y by D ⁇ 1 y, and uses a DNN learned in advance from a vector in which the absolute value of each component is taken to obtain x. Estimate the complexity of ⁇ , that is, the degree of sparsity K ⁇ . The sparseness estimation unit 21 sends the estimated sparseness K ⁇ to the parameter determination unit 22.
  • the DNN can be composed of two or three hidden layers.
  • the frame length is L and a DFT matrix is used as D
  • the first layer has L/2 dimensions.
  • the training data used for this DNN pre-learning consists of input data given to the DNN input layer and teacher data given to the DNN output layer. Each data is generated as follows. • Collect samples of the expected input signal x. -The signal expression vector of the signal y after clipping is found by D -1 y, and the vector that takes the absolute value of each component is used as the input data. -The signal expression vector of the signal x before clipping is calculated by D -1 x, the absolute value square of each element is calculated, and sorted in descending order. The sum is calculated from the larger elements, and the number of elements is calculated when the sum exceeds 95% of the total. This number of elements is defined as the degree of sparseness of the signal and used as teacher data.
  • step S22 the parameter determination unit 22 corrects the parameters s and max_iter for SPADE from the estimated sparse degree K ⁇ using the equations (3) and (4).
  • the parameter determination unit 22 sends the modified parameters s_rev and max_iter_rev to the waveform restoration unit 12.
  • step S12 the waveform restoration unit 12 estimates the signal frame x before clipping by executing the SPADE process using the corrected update amount s_rev and the corrected maximum number of repetitions max_iter_rev.
  • the waveform restoration unit 12 sends the estimated signal frame x before clipping to the frame synthesis unit 13.
  • step S13 the frame synthesis unit 13 applies frame synthesis processing to the estimated signal frame x before clipping to restore the signal before clipping.
  • the program describing this processing content can be recorded in a computer-readable recording medium.
  • the computer-readable recording medium may be, for example, a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, or the like.
  • distribution of this program is performed by selling, transferring, or lending a portable recording medium such as a DVD or a CD-ROM in which the program is recorded.
  • the program may be stored in a storage device of a server computer and transferred from the server computer to another computer via a network to distribute the program.
  • a computer that executes such a program first stores, for example, the program recorded in a portable recording medium or the program transferred from the server computer in its own storage device. Then, when executing the process, this computer reads the program stored in its own storage device and executes the process according to the read program.
  • a computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to this computer. Each time, the processing according to the received program may be sequentially executed.
  • the above-mentioned processing is executed by the so-called ASP (Application Service Provider) type service that realizes the processing function only by executing the execution instruction and the result acquisition without transferring the program from the server computer to this computer.
  • ASP Application Service Provider
  • the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (such as data that is not a direct command to a computer but has the property of defining computer processing).
  • the device is configured by executing a predetermined program on a computer, but at least a part of the processing content may be implemented by hardware.

Landscapes

  • Engineering & Computer Science (AREA)
  • Human Computer Interaction (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Computational Linguistics (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Quality & Reliability (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Circuit For Audible Band Transducer (AREA)
  • Auxiliary Devices For Music (AREA)

Abstract

A waveform restoration process is reliably performed even in conditions in which the usable calculation amount is limited. This waveform restoration device 10 repeatedly estimates, from a post-clip signal, a pre-clip signal having a prescribed degree of complexity, while increasing the degree of complexity by a prescribed update amount, thereby restoring the pre-clip signal. A parameter determination unit 12 determines a maximum number of repetitions on the basis of the post-clip signal. A waveform restoration unit 13 repeatedly estimates the pre-clip signal using the maximum number of repetitions as the maximum number for the estimate.

Description

復元装置、復元方法、およびプログラムRestoration apparatus, restoration method, and program
　この発明は、クリップ後の信号からクリップ前の信号を復元する技術に関する。 The present invention relates to a technique for restoring a signal before clipping from a signal after clipping.
　装置間で信号を入出力する際に、信号振幅が装置の入出力レンジより大きい部分は一定値にクリッピングされる。クリッピングは、例えば、センサーから信号を得るとき、信号を何らかの機器に出力するとき、アナログ信号をA/Dコンバータに入力してデジタル化するとき等、広範囲な状況で生じる可能性がある。そこで、クリップされた信号からクリップ前の信号波形を復元する研究がこれまでなされてきた。 When inputting/outputting a signal between devices, the part where the signal amplitude is larger than the input/output range of the device is clipped to a constant value. Clipping can occur in a wide range of situations, for example, when obtaining a signal from a sensor, when outputting a signal to some device, when inputting an analog signal into an A/D converter for digitization, and the like. Therefore, studies have been made to restore the signal waveform before clipping from the clipped signal.
　そのような方法として、SPADE（SParse Audio DEclipper）と呼ばれる方法が提案されている（非特許文献１）。以下、SPADEについて説明する。 As such a method, a method called SPADE (SParse Audio DEclipper) has been proposed (Non-Patent Document 1). The SPADE will be described below.
　なお、文中で使用する記号「－」「^」は、本来直前の文字の真上に記載されるべきものであるが、テキスト記法の制限により、当該文字の直後に記載する。数式中においてはこれらの記号は本来の位置、すなわち文字の真上に記述している。例えば、「z-」は数式中では次式で表される。
Figure JPOXMLDOC01-appb-M000001
 また、例えば、「z^」は数式中では次式で表される。
Figure JPOXMLDOC01-appb-M000002
 The symbols " - " and "^" used in the sentence should be written right above the character just before, but due to the limitation of text notation, they are written immediately after the character. In the mathematical formula, these symbols are described at their original positions, that is, directly above the characters. For example, "z - " is represented by the following equation in the mathematical formula.
Figure JPOXMLDOC01-appb-M000001
 Further, for example, “z^” is represented by the following equation in the mathematical formula.
Figure JPOXMLDOC01-appb-M000002
　本来の信号（クリップ前の信号）を信号ベクトルx=[x1, …, xN]、クリップされた信号を信号ベクトルy=[y1, …, yN]で表現する。クリッピング前後の信号の各サンプルは、式（１）の関係がある。
Figure JPOXMLDOC01-appb-M000003
 The original signal (the signal before clipping) is represented by a signal vector x=[x 1 ,..., X N ], and the clipped signal is represented by a signal vector y=[y 1 ,..., Y N ]. Each sample of the signal before and after clipping has the relationship of Formula (1).
Figure JPOXMLDOC01-appb-M000003
　ここでθはクリッピングレベルである。クリッピング後の信号サンプルは、上限でクリップされる信号サンプルS+、クリップされない信号サンプルSr、下限でクリップされる信号サンプルS-のいずれかに属する。 Where θ is the clipping level. The signal sample after clipping belongs to one of the signal sample S + that is clipped at the upper limit, the signal sample S r that is not clipped, and the signal sample S − that is clipped at the lower limit.
　SPADEでは、まず辞書行列Dを定める。そして、辞書行列Dの逆行列D-1を信号ベクトルxにかけて得られる信号表現ベクトルzに注目して、信号の複雑さをz中の非ゼロ要素の数、すなわちzのL0ノルム||z||0ではかる。辞書行列Dとしては、DFT行列（Discrete Fourier Transform Matrix）やDCT行列（Discrete Cosine Transform Matrix）等が用いられる。 In SPADE, the dictionary matrix D is first defined. Then, paying attention to the signal expression vector z obtained by multiplying the inverse matrix D −1 of the dictionary matrix D by the signal vector x, the complexity of the signal is determined by the number of non-zero elements in z, that is, the L 0 norm of z ||z || Measures with 0 . As the dictionary matrix D, a DFT matrix (Discrete Fourier Transform Matrix), a DCT matrix (Discrete Cosine Transform Matrix), or the like is used.
　SPADEは、クリップ前の信号の複雑さ（以下、「複雑度」とも呼ぶ）をkとし、所定の更新量sを複雑さkの初期値として想定する。まず、入力信号すなわちクリップ後の信号yを、D-1で信号表現ベクトルzへ変換する。z中で絶対値の大きい方からk個の要素を残し、それ以外の値を0にすることで、複雑さkの信号表現ベクトルz-へと変換する。この操作はhard thresholdingと呼ばれ、数式でz-=Hk(z)と表現される（下記表１のステップ２に対応）。次に、この信号表現ベクトルz-にDをかけて、推定信号ベクトルx-=Dz-へ変換する。推定信号ベクトルx-は、この段階でのクリップ前の信号ベクトルxの推定結果になる。通常、この推定信号ベクトルx-と入力信号ベクトルyでは非クリップ部分でも乖離がある。そこで下記の二つの条件を満たす信号表現ベクトルz^を求める（下記表１のステップ３に対応）。
　条件１．クリップされたDz^がyと一致する。
　条件２．z^とz-の距離が最小になる。 SPADE assumes that the complexity of a signal before clipping (hereinafter, also referred to as “complexity”) is k, and a predetermined update amount s is an initial value of the complexity k. First, the input signal, that is, the signal y after clipping is converted into a signal expression vector z at D −1 . leaving k elements in z from the larger absolute value, the other value by zero, signal representation vector z complexity k - converted to. This operation is called hard thresholding, and is expressed as z − =H k (z) in the mathematical formula (corresponding to step 2 in Table 1 below). Then, the signal representation vector z - over D, the estimated signal vector x - = Dz - converting into. The estimated signal vector x - is the estimation result of the signal vector x before clipping at this stage. Usually, there is a discrepancy between the estimated signal vector x - and the input signal vector y even in the non-clip part. Therefore, a signal expression vector z^ satisfying the following two conditions is obtained (corresponding to step 3 in Table 1 below).
Condition 1. The clipped Dz^ matches y.
Condition 2. z ^ and z - distance is the minimum.
　z^とz-の距離があらかじめ決めた値より大きい場合には、「想定する信号の複雑さkが不足しているためにターゲット信号を表現できない」と判定して複雑さkを更新量sずつ増やし、上記の処理を繰り返す。 z ^ and z - if the value is greater than the distance is determined in advance of, "can not express the target signal in order to complexity k of the expected signal is insufficient," the decision to the complexity k update amount s And the above process is repeated.
　以上の処理を、最適化手法ADMM（非特許文献２）を用いて実装すると、表１のアルゴリズムが得られる。 By implementing the above processing using the optimization method ADMM (Non-Patent Document 2), the algorithm shown in Table 1 is obtained.
Figure JPOXMLDOC01-appb-T000004
Figure JPOXMLDOC01-appb-T000004
　なお、表１では、複雑さkを固定してz-からz^を推定する処理をr回行っている。通常r=1に設定される。また、max_iterは上記のz-からz^を推定する処理の最大繰り返し回数であり、式（２）で与えられる。
Figure JPOXMLDOC01-appb-M000005
 ただしceil()は切り上げ関数であり、dim(z)はベクトルzの次元である。 In Table 1, to fix the complexity k z - is the process of estimating the z ^ after performing r times. Normally set to r=1. Further, max_iter the above z - the maximum number of repetitions of the process of estimating the z ^ from, given by equation (2).
Figure JPOXMLDOC01-appb-M000005
 Where ceil() is the round-up function and dim(z) is the dimension of vector z.
　SPADEは、通常のフレーム信号処理と組み合わせて使用される。つまり、図１に示すように、入力されたクリップ後の信号は、フレーム分割部１１で重なりをもつ一定長のフレームに分割され、各フレームに窓かけ処理をした後、波形復元部１２で上記SPADE処理が適用される。そして、フレーム合成部１３でその処理結果にフレーム合成処理が適用され、復元したクリップ前の信号が得られる。 -SPADE is used in combination with normal frame signal processing. That is, as shown in FIG. 1, the input signal after clipping is divided into frames of a certain length with overlapping in the frame dividing unit 11, and after windowing processing is performed on each frame, the waveform restoring unit 12 SPADE processing is applied. Then, the frame synthesizing unit 13 applies the frame synthesizing process to the processing result, and the restored signal before the clip is obtained.
　なお、非特許文献３によれば、表１のステップ３のz-からz^を推定する処理は、射影
Figure JPOXMLDOC01-appb-M000006
 を用い、
Figure JPOXMLDOC01-appb-M000007
 で実現される。 Incidentally, according to Non-Patent Document 3, the z Step 3 of Table 1 - process of estimating the z ^ from the projection
Figure JPOXMLDOC01-appb-M000006
 Using
Figure JPOXMLDOC01-appb-M000007
 Will be realized in.
　しかしながら、例えば、IoT（Internet of Thing）等においてリアルタイムに複数のセンサー信号の波形を復元する必要がある場合には、SPADEの演算量は大きいという問題がある。それは、SPADEが複雑さの想定kを順次増やしながら波形復元処理を進める必要があるためであり、また、そもそも入力信号の複雑さが未知で絶えず変動しているためである。 However, for example, when it is necessary to restore the waveforms of multiple sensor signals in real time in IoT (Internet of Things), there is a problem that the calculation amount of SPADE is large. This is because SPADE needs to proceed with the waveform restoration process while sequentially increasing the assumed k of complexity, and the complexity of the input signal is unknown and constantly fluctuates.
　この発明の目的は、上記のような技術的課題を鑑みて、使用可能な演算量が制限されている状況であっても確実に波形復元を行うことができる技術を実現することである。 The object of the present invention is to realize a technique capable of surely performing waveform restoration even in a situation where the usable calculation amount is limited, in view of the above technical problems.
　上記の課題を解決するために、この発明の一態様の復元装置は、クリップ後信号から所定の複雑度のクリップ前信号を推定することを、複雑度を所定の更新量だけ増加させながら繰り返し行うことで、クリップ前信号を復元する復元装置であって、クリップ後信号に基づいて最大繰り返し数を決定するパラメータ決定部と、最大繰り返し数を推定の最大回数としてクリップ前信号の推定を繰り返し行う推定部と、を含む。 In order to solve the above problems, the restoration device according to one aspect of the present invention repeatedly estimates a pre-clip signal of a predetermined complexity from a clipped signal while increasing the complexity by a predetermined update amount. Thus, a restoration device for restoring the pre-clip signal, a parameter determination unit that determines the maximum number of repetitions based on the post-clip signal, and an estimation that repeatedly estimates the pre-clip signal with the maximum number of repetitions as the maximum number of estimations And a part.
　この発明の波形復元技術によれば、使用可能な演算量が制限されている状況であっても波形復元処理を確実に行うことが可能となる。 According to the waveform restoration technology of the present invention, it becomes possible to reliably perform the waveform restoration process even in a situation where the usable calculation amount is limited.
図１は、従来の波形復元装置の機能構成を例示する図である。FIG. 1 is a diagram illustrating a functional configuration of a conventional waveform restoration device. 図２は、実施形態の波形復元装置の機能構成を例示する図である。FIG. 2 is a diagram illustrating a functional configuration of the waveform restoration device according to the embodiment. 図３は、実施形態の波形復元方法の処理手順を例示する図である。FIG. 3 is a diagram illustrating a processing procedure of the waveform restoration method of the embodiment.
　以下、この発明の実施の形態について詳細に説明する。なお、図面中において同じ機能を有する構成部には同じ番号を付し、重複説明を省略する。 Hereinafter, embodiments of the present invention will be described in detail. In the drawings, components having the same function are denoted by the same reference numerals, and duplicate description will be omitted.
　［第一実施形態］
　第一実施形態の波形復元装置は、まず、ディープニューラルネットワーク（DNN: Deep Neural Network）（参考文献１参照）を用いて、クリップ後の信号からクリップ前の信号の複雑さ（スパース度）を推定する。また、推定処理の最大繰り返し数を演算量の制限指標としてあらかじめ指定し、推定されたスパース度に基づいて複雑さkの更新量sを制御する。これにより、使用可能な演算量が制限されている状況であっても波形復元処理を確実に行うことを可能とする。
　〔参考文献１〕I. Goodfellow, Y. Bensio, and A. Courville, "Deep learning", MIT Press, 2016. [First embodiment]
The waveform restoration device of the first embodiment first estimates the complexity (sparseness) of a signal before clipping from a signal after clipping using a deep neural network (DNN: Deep Neural Network) (see Reference Document 1). To do. In addition, the maximum number of iterations of the estimation process is designated in advance as a limit index of the amount of calculation, and the update amount s of complexity k is controlled based on the estimated degree of sparsity. This makes it possible to reliably perform the waveform restoration process even in a situation where the amount of calculation that can be used is limited.
[Reference 1] I. Goodfellow, Y. Bensio, and A. Courville, "Deep learning", MIT Press, 2016.
　なお、許容する推定処理の最大繰り返し数は、本来必要なmax_iterのα倍（0<α<1）で指定されるものとする。 Note that the maximum number of iterations of the estimation process that is allowed is specified by α times (0<α<1) of max_iter which is originally required.
　第一実施形態の波形復元装置１０（以下、「復元装置」とも呼ぶ）は、図２に例示するように、フレーム分割部１１、波形復元部１２（以下、「推定部」とも呼ぶ）、およびフレーム合成部１３に加えて、スパース度推定部２１およびパラメータ決定部２２を備える。この波形復元装置１０が、図３に例示する各ステップの処理を行うことにより第一実施形態の波形復元方法が実現される。 The waveform restoration device 10 (hereinafter, also referred to as “restoration device”) of the first embodiment, as illustrated in FIG. 2, a frame division unit 11, a waveform restoration unit 12 (hereinafter, also referred to as “estimation unit”), and In addition to the frame synthesis section 13, a sparse degree estimation section 21 and a parameter determination section 22 are provided. The waveform restoration apparatus 10 implements the waveform restoration method of the first embodiment by performing the processing of each step illustrated in FIG.
　波形復元装置１０は、例えば、中央演算処理装置（CPU: Central Processing Unit）、主記憶装置（RAM: Random Access Memory）などを有する公知又は専用のコンピュータに特別なプログラムが読み込まれて構成された特別な装置である。波形復元装置１０は、例えば、中央演算処理装置の制御のもとで各処理を実行する。波形復元装置１０に入力されたデータや各処理で得られたデータは、例えば、主記憶装置に格納され、主記憶装置に格納されたデータは必要に応じて中央演算処理装置へ読み出されて他の処理に利用される。波形復元装置１０の各処理部は、少なくとも一部が集積回路等のハードウェアによって構成されていてもよい。 The waveform restoration device 10 is, for example, a special program configured by reading a special program into a known or dedicated computer having a central processing unit (CPU: Central Processing Unit), a main storage device (RAM: Random Access Memory), and the like. It is a device. The waveform restoration device 10 executes each process under the control of the central processing unit, for example. The data input to the waveform restoration device 10 and the data obtained by each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out to the central processing unit as necessary. It is used for other processing. At least a part of each processing unit of the waveform restoration device 10 may be configured by hardware such as an integrated circuit.
　ステップＳ１１において、フレーム分割部１１は、入力されたクリップ後の信号をフレーム分割する。フレーム分割部１１は、クリップ後の信号フレームyを波形復元部１２およびスパース度推定部２１へ送る。 In step S11, the frame dividing unit 11 divides the input clipped signal into frames. The frame division unit 11 sends the clipped signal frame y to the waveform restoration unit 12 and the sparsity degree estimation unit 21.
　ステップＳ２１において、スパース度推定部２１は、クリップ後の信号フレームyの信号表現ベクトルをD-1yにより求め、その各成分の絶対値をとったベクトルから、あらかじめ学習したDNNを用いて、x^の複雑さ、すなわちスパース度K^を推定する。スパース度推定部２１は、推定したスパース度K^をパラメータ決定部２２へ送る。 In step S21, the sparseness estimation unit 21 obtains a signal expression vector of the clipped signal frame y by D −1 y, and uses a DNN learned in advance from a vector in which the absolute value of each component is taken to obtain x. Estimate the complexity of ^, that is, the degree of sparsity K^. The sparseness estimation unit 21 sends the estimated sparseness K^ to the parameter determination unit 22.
　このDNNは、隠れ層が２ないし３層のDNNで構成可能である。フレーム長がLで、DとしてDFT行列を用いた場合、１層目はL/2次元になる。このDNNの事前学習に使うトレーニングデータは、DNN入力層に与える入力データとDNN出力層に与える教師データからなる。各データは以下のように生成する。
　・想定される入力信号xのサンプルを集める。
　・クリップ後の信号yの信号表現ベクトルをD-1yにより求め、その各成分の絶対値をとったベクトルを入力データとする。
　・クリップ前の信号xの信号表現ベクトルをD-1xにより求め、各要素の絶対値２乗を求め、大きい順にソートする。大きい要素から和をとり、その和が総和の95％を越える際の要素数を求める。この要素数を信号のスパース度と定義し、教師データとする。 The DNN can be composed of two or three hidden layers. When the frame length is L and a DFT matrix is used as D, the first layer has L/2 dimensions. The training data used for this DNN pre-learning consists of input data given to the DNN input layer and teacher data given to the DNN output layer. Each data is generated as follows.
• Collect samples of the expected input signal x.
-The signal expression vector of the signal y after clipping is found by D -1 y, and the vector that takes the absolute value of each component is used as the input data.
-The signal expression vector of the signal x before clipping is calculated by D -1 x, the absolute value square of each element is calculated, and sorted in descending order. The sum is calculated from the larger elements, and the number of elements is calculated when the sum exceeds 95% of the total. This number of elements is defined as the degree of sparseness of the signal and used as teacher data.
　ステップＳ２２において、パラメータ決定部２２は、推定されたスパース度K^から式（３）（４）を用いて、SPADE向けのパラメータsとmax_iterとを修正する。パラメータ決定部２２は、修正されたパラメータs_revとmax_iter_revとを波形復元部１２へ送る。
Figure JPOXMLDOC01-appb-M000008
 In step S22, the parameter determination unit 22 corrects the parameters s and max_iter for SPADE from the estimated sparse degree K^ using the equations (3) and (4). The parameter determination unit 22 sends the modified parameters s_rev and max_iter_rev to the waveform restoration unit 12.
Figure JPOXMLDOC01-appb-M000008
　ステップＳ１２において、波形復元部１２は、修正された更新量s_revおよび修正された最大繰り返し数max_iter_revを用いてSPADE処理を実行することで、クリップ前の信号フレームxを推定する。波形復元部１２は、推定したクリップ前の信号フレームxをフレーム合成部１３へ送る。 In step S12, the waveform restoration unit 12 estimates the signal frame x before clipping by executing the SPADE process using the corrected update amount s_rev and the corrected maximum number of repetitions max_iter_rev. The waveform restoration unit 12 sends the estimated signal frame x before clipping to the frame synthesis unit 13.
　ステップＳ１３において、フレーム合成部１３は、推定されたクリップ前の信号フレームxにフレーム合成処理を適用し、クリップ前の信号を復元する。 In step S13, the frame synthesis unit 13 applies frame synthesis processing to the estimated signal frame x before clipping to restore the signal before clipping.
　［第二実施形態］
　SPADE内部において、繰り返し数iに応じて複雑さkの更新量sを増やす制御を行う。そのような制御の一例として、更新量sを一定間隔でT段階に切り替えることが考えられる。このケースではCを式（５）を満たす最大の自然数とし、
Figure JPOXMLDOC01-appb-M000009
 複雑さkの更新量sを式（６）に示す繰り返し数iの関数s(i)で決める。
Figure JPOXMLDOC01-appb-M000010
 ただしfloor()は切り下げ関数である。 [Second embodiment]
Within SPADE, control is performed to increase the update amount s of complexity k according to the number of iterations i. As an example of such control, it is conceivable to switch the update amount s to T stages at regular intervals. In this case, let C be the maximum natural number that satisfies equation (5),
Figure JPOXMLDOC01-appb-M000009
 The update amount s of the complexity k is determined by the function s(i) of the iteration number i shown in the equation (6).
Figure JPOXMLDOC01-appb-M000010
 However, floor() is a round-down function.
　s(i)は、SPADE内でz-からz^の推定をC回繰り返す度に、1, 2, …, 2Tと増えていく。この制御では、SPADEの繰り返し初期では、複雑さkをゆっくり増加させながらz-からz^の推定を行い、その後、複雑さkの増加を大きくしなからz-からz^の推定を行う。 s(i) increases as 1, 2, …, 2 T each time the estimation of z − to z^ is repeated C times in SPADE. In this control, the repetitive initial SPADE, while increasing slowly complexity k z - performs estimation from z ^ of, then the complexity k of increase was significantly Nakara z - performing the z ^ estimation.
　SPADEでは、複雑さkの更新量sを大きく設定すると復元後の信号の振幅が小さくなる傾向がある。そのため繰り返し初期で複雑さkの増加をゆっくりにすることで、復元後の信号の推定精度を上げることができる。 In SPADE, if the update amount s of complexity k is set large, the amplitude of the restored signal tends to become small. Therefore, the estimation accuracy of the restored signal can be improved by slowing the increase of the complexity k at the initial stage of iteration.
　［変形例］
　C2を式（７）を満たす最大の自然数とし、
Figure JPOXMLDOC01-appb-M000011
 複雑さkの更新量sを式（８）に示す関数s(i)で決める。
Figure JPOXMLDOC01-appb-M000012
 [Modification]
Let C 2 be the maximum natural number that satisfies equation (7),
Figure JPOXMLDOC01-appb-M000011
 The update amount s of complexity k is determined by the function s(i) shown in equation (8).
Figure JPOXMLDOC01-appb-M000012
　以上、この発明の実施の形態について説明したが、具体的な構成は、これらの実施の形態に限られるものではなく、この発明の趣旨を逸脱しない範囲で適宜設計の変更等があっても、この発明に含まれることはいうまでもない。実施の形態において説明した各種の処理は、記載の順に従って時系列に実行されるのみならず、処理を実行する装置の処理能力あるいは必要に応じて並列的にあるいは個別に実行されてもよい。 Although the embodiments of the present invention have been described above, the specific configuration is not limited to these embodiments, and even if the design is appropriately changed without departing from the gist of the present invention, Needless to say, it is included in the present invention. The various kinds of processing described in the embodiments may be executed not only in time series according to the order described, but also in parallel or individually according to the processing capability of the device that executes the processing or the need.
　［プログラム、記録媒体］
　上記実施形態で説明した各装置における各種の処理機能をコンピュータによって実現する場合、各装置が有すべき機能の処理内容はプログラムによって記述される。そして、このプログラムをコンピュータで実行することにより、上記各装置における各種の処理機能がコンピュータ上で実現される。 [Program, recording medium]
When various processing functions in each device described in the above embodiment are realized by a computer, processing contents of functions that each device should have are described by a program. By executing this program on a computer, various processing functions of the above-described devices are realized on the computer.
　この処理内容を記述したプログラムは、コンピュータで読み取り可能な記録媒体に記録しておくことができる。コンピュータで読み取り可能な記録媒体としては、例えば、磁気記録装置、光ディスク、光磁気記録媒体、半導体メモリ等どのようなものでもよい。 The program describing this processing content can be recorded in a computer-readable recording medium. The computer-readable recording medium may be, for example, a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, or the like.
　また、このプログラムの流通は、例えば、そのプログラムを記録したDVD、CD-ROM等の可搬型記録媒体を販売、譲渡、貸与等することによって行う。さらに、このプログラムをサーバコンピュータの記憶装置に格納しておき、ネットワークを介して、サーバコンピュータから他のコンピュータにそのプログラムを転送することにより、このプログラムを流通させる構成としてもよい。 Also, distribution of this program is performed by selling, transferring, or lending a portable recording medium such as a DVD or a CD-ROM in which the program is recorded. Further, the program may be stored in a storage device of a server computer and transferred from the server computer to another computer via a network to distribute the program.
　このようなプログラムを実行するコンピュータは、例えば、まず、可搬型記録媒体に記録されたプログラムもしくはサーバコンピュータから転送されたプログラムを、一旦、自己の記憶装置に格納する。そして、処理の実行時、このコンピュータは、自己の記憶装置に格納されたプログラムを読み取り、読み取ったプログラムに従った処理を実行する。また、このプログラムの別の実行形態として、コンピュータが可搬型記録媒体から直接プログラムを読み取り、そのプログラムに従った処理を実行することとしてもよく、さらに、このコンピュータにサーバコンピュータからプログラムが転送されるたびに、逐次、受け取ったプログラムに従った処理を実行することとしてもよい。また、サーバコンピュータから、このコンピュータへのプログラムの転送は行わず、その実行指示と結果取得のみによって処理機能を実現する、いわゆるASP（Application Service Provider）型のサービスによって、上述の処理を実行する構成としてもよい。なお、本形態におけるプログラムには、電子計算機による処理の用に供する情報であってプログラムに準ずるもの（コンピュータに対する直接の指令ではないがコンピュータの処理を規定する性質を有するデータ等）を含むものとする。 A computer that executes such a program first stores, for example, the program recorded in a portable recording medium or the program transferred from the server computer in its own storage device. Then, when executing the process, this computer reads the program stored in its own storage device and executes the process according to the read program. As another execution form of this program, a computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to this computer. Each time, the processing according to the received program may be sequentially executed. In addition, the above-mentioned processing is executed by the so-called ASP (Application Service Provider) type service that realizes the processing function only by executing the execution instruction and the result acquisition without transferring the program from the server computer to this computer. May be Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (such as data that is not a direct command to a computer but has the property of defining computer processing).
　また、この形態では、コンピュータ上で所定のプログラムを実行させることにより、本装置を構成することとしたが、これらの処理内容の少なくとも一部をハードウェア的に実現することとしてもよい。 In this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing content may be implemented by hardware.
１０、９０　波形復元装置
１１　フレーム分割部
１２　波形復元部
１３　フレーム合成部
２１　スパース度推定部
２２　パラメータ決定部 10, 90 waveform restoration device 11 frame division unit 12 waveform restoration unit 13 frame synthesis unit 21 sparseness estimation unit 22 parameter determination unit

Claims (5)

  1. 　クリップ後信号から所定の複雑度のクリップ前信号を推定することを、上記複雑度を所定の更新量だけ増加させながら繰り返し行うことで、クリップ前信号を復元する復元装置であって、
    　上記クリップ後信号に基づいて最大繰り返し数を決定するパラメータ決定部と、
    　上記最大繰り返し数を上記推定の最大回数として上記クリップ前信号の推定を繰り返し行う推定部と、
    　を含む復元装置。     A restoration device that restores a pre-clip signal by repeatedly estimating a pre-clip signal of a predetermined complexity from a post-clip signal while increasing the complexity by a predetermined update amount,
    A parameter determination unit that determines the maximum number of repetitions based on the clipped signal,
    An estimation unit that repeatedly estimates the pre-clip signal with the maximum number of repetitions as the maximum number of times of the estimation,
    Restoration device including.
  2. 　請求項１に記載の復元装置であって、
    　上記クリップ後信号から上記クリップ前信号のスパース度を推定するスパース度推定部をさらに含み、
    　上記パラメータ決定部は、上記クリップ前信号のスパース度に基づいてあらかじめ定めた上記最大繰り返し数および上記更新量を修正する、
    　復元装置。     The restoration device according to claim 1, wherein
    Further comprising a sparseness estimation unit that estimates the sparseness of the pre-clip signal from the clipped signal,
    The parameter determination unit corrects the maximum number of repetitions and the update amount that are predetermined based on the sparsity of the pre-clip signal,
    Restoration device.
  3. 　請求項１または２に記載の復元装置であって、
    　上記推定部は、上記推定を繰り返した回数に応じて上記更新量を増加させる、
    　復元装置。     The restoration device according to claim 1 or 2, wherein
    The estimation unit increases the update amount according to the number of times the estimation is repeated,
    Restoration device.
  4. 　クリップ後信号から所定の複雑度のクリップ前信号を推定することを、上記複雑度を所定の更新量だけ増加させながら繰り返し行うことで、クリップ前信号を復元する復元方法であって、
    　パラメータ決定部が、上記クリップ後信号に基づいて最大繰り返し数を決定し、
    　推定部が、上記最大繰り返し数を上記推定の最大回数として上記クリップ前信号の推定を繰り返し行う、
    　復元方法。     A restoration method for restoring a pre-clip signal by repeatedly estimating the pre-clip signal of a predetermined complexity from the post-clip signal while increasing the complexity by a predetermined update amount,
    The parameter determination unit determines the maximum number of repetitions based on the clipped signal,
    The estimation unit repeatedly estimates the pre-clip signal with the maximum number of repetitions as the maximum number of times of estimation,
    How to restore.
  5. 　請求項１から３のいずれかに記載の復元装置としてコンピュータを機能させるためのプログラム。 A program for causing a computer to function as the restoration device according to any one of claims 1 to 3.
PCT/JP2019/049084 2018-12-28 2019-12-16 Restoration device, restoration method, and program WO2020137641A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2018246762A JP2020106713A (en) 2018-12-28 2018-12-28 Restoring device, restoring method, and program
JP2018-246762 2018-12-28

Publications (1)

Publication Number Publication Date
WO2020137641A1 true WO2020137641A1 (en) 2020-07-02

Family

ID=71129118

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2019/049084 WO2020137641A1 (en) 2018-12-28 2019-12-16 Restoration device, restoration method, and program

Country Status (2)

Country Link
JP (1) JP2020106713A (en)
WO (1) WO2020137641A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022070315A1 (en) * 2020-09-30 2022-04-07 日本電信電話株式会社 Waveform restoration device, waveform restoration method, and program

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5987407A (en) * 1997-10-28 1999-11-16 America Online, Inc. Soft-clipping postprocessor scaling decoded audio signal frame saturation regions to approximate original waveform shape and maintain continuity
US20110122930A1 (en) * 2009-11-25 2011-05-26 Al-Naffouri Tareq Y OFDM peak-to-average power ratio reduction method
US20130129115A1 (en) * 2009-02-26 2013-05-23 Paris Smaragdis System and Method for Dynamic Range Extension Using Interleaved Gains
US20130177089A1 (en) * 2012-01-09 2013-07-11 King Fahd University Of Petroleum And Minerals Method of performing peak reduction and clipping mitigation

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5987407A (en) * 1997-10-28 1999-11-16 America Online, Inc. Soft-clipping postprocessor scaling decoded audio signal frame saturation regions to approximate original waveform shape and maintain continuity
US20130129115A1 (en) * 2009-02-26 2013-05-23 Paris Smaragdis System and Method for Dynamic Range Extension Using Interleaved Gains
US20110122930A1 (en) * 2009-11-25 2011-05-26 Al-Naffouri Tareq Y OFDM peak-to-average power ratio reduction method
US20130177089A1 (en) * 2012-01-09 2013-07-11 King Fahd University Of Petroleum And Minerals Method of performing peak reduction and clipping mitigation

Also Published As

Publication number Publication date
JP2020106713A (en) 2020-07-09

Similar Documents

Publication Publication Date Title
Blanchard et al. GPU accelerated greedy algorithms for compressed sensing
TW202119293A (en) Method and system of quantizing artificial neural network and arti ficial neural network apparatus
JP2021504836A5 (en)
US20190065443A1 (en) Matrix generation program, matrix generation apparatus, and plagiarism detection program
US20210326756A1 (en) Methods of providing trained hyperdimensional machine learning models having classes with reduced elements and related computing systems
US10262680B2 (en) Variable sound decomposition masks
JP2017049907A (en) Neural network, learning method therefor and computer program
Llanas et al. Constructive approximation of discontinuous functions by neural networks
WO2020137641A1 (en) Restoration device, restoration method, and program
Kolountzakis Fourier pairs of discrete support with little structure
Kulesh et al. Adaptive metrics in the nearest neighbours method
US20210174815A1 (en) Quantization method of latent vector for audio encoding and computing device for performing the method
WO2022070315A1 (en) Waveform restoration device, waveform restoration method, and program
JPWO2020031281A1 (en) Information processing equipment, information processing methods, and programs
JP2019074621A (en) Signal separation device, signal separation method and program
Li et al. Dictionary learning with the ℓ _ 1/2 ℓ 1/2-regularizer and the coherence penalty and its convergence analysis
Addison et al. Blind source separation with non-stationary mixing using wavelets
CN114332561A (en) Super-resolution model training method, device, equipment and medium
Cyrus et al. Meta-interpretive Learning from Fractal Images
Ahmedov et al. Numerical integration of the integrals based on Haar wavelets
Hasenclever et al. Variational inference with orthogonal normalizing flows
JP7188589B2 (en) Restoration device, restoration method and program
Ojobor A Collocation method for second order boundary value problems
Gadzhiev On the variance of a centered random value roundoff error
JP6994572B2 (en) Data processing system and data processing method

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19905763

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19905763

Country of ref document: EP

Kind code of ref document: A1