SYSTEM AND METHOD FOR PROCESSING AN AUDIO INPUT SIGNAL

Abstract

System (100) for processing an audio input signal (15), wherein the system (100) comprises: a microphone (20), a controller (10), a data storage device (50) and a communication link (60) with a remote audio speaker (70); wherein the microphone (20) is configured to capture and generate the audio input signal (15) and to transmit the audio input signal (15) to the controller (10); wherein the controller (10) is functionally connected to the communication link (60); and wherein the data storage (50) contains instructions that can be executed by the controller (10), wherein the instructions contain: Generating a first resultant (235) based on the audio input signal (15) via a linear noise reduction filtering algorithm; Generating a second resultant (245) based on the first resultant (235) using a nonlinear post-filtering algorithm; Generating an audio output signal based on the second resultant (245) via a feature recovery algorithm (300); and Transmitting the audio output signal to the remote audio speaker (70) via the communication link (60); characterized by the fact that the feature recovery algorithm (300) comprises a module based on a deep neural network (DNN-based module) which includes: a STFT (short-term Fourier transform) (310); several convolutional layers (320); a first LSTM (long-short-term memory layer) (330); a second LSTM (332); a dense layer (340); several transposed convolutional layers (350); and an inverse STFT (ISTFT) (370).

Description

[0001] The present invention relates to a system according to the preamble of claim 1 and a method according to the preamble of claim 9 for processing an audio input signal, as is substantially known from US 2019 / 0 122 689 A1. INTRODUCTION

[0002] Speech processing systems encompass the use of speakerphone-type devices such as smartphones, video conferencing systems, laptops, and tablets. In some systems, the speaker may be located in an enclosed space and at a relatively large distance from the microphone. Such arrangements can introduce ambient noise, including background noise, interference, and reverberation. These arrangements can lead to acoustic signal processing challenges that negatively impact sound quality and the associated signal-to-noise ratio (SNR).

[0003] Speech processing technologies such as automatic speech recognition (ASR) and teleconferencing often incorporate noise reduction strategies and systems to lower audible background noise and improve speech intelligibility. Noise reduction systems can include linear noise reduction algorithms, nonlinear post-filtering algorithms, and others. The performance of linear noise reduction algorithms may not be sufficient to achieve a desired signal-to-noise ratio (SNR) target.

[0004] A nonlinear post-filtering (PF) algorithm, connected in series with a linear noise reduction algorithm, can improve noise reduction levels, but there are trade-offs between residual noise and speech distortion levels. Removing speech characteristics from the signal due to spectral subtraction algorithms, which can be used in a PF module, can introduce sound distortion. Achieving a target signal-to-noise ratio (SNR) with minimal speech distortion requires precise tuning of such a system, which can be challenging.

[0005] Therefore, there is a need for an improved procedure and system for speech processing that includes noise reduction strategies to decrease audible background noise levels, improve speech intelligibility, and reduce the need for precise tuning. SUMMARY

[0006] According to the invention, a system with the features of claim 1 and a method with the features of claim 9 for processing an audio input signal are presented.

[0007] The concepts described here create methods and systems for speech processing that include noise reduction strategies to reduce audible background noise levels and improve speech intelligibility.

[0008] The concepts include a system for processing an audio input signal, utilizing a microphone, a controller, and a communication link that can be coupled to a remote audio speaker. The microphone is configured to capture and generate the audio input signal and transmit it to the controller, which is coupled to the communication link. The controller contains executable code for generating a first resultant based on the audio input signal using a linear noise reduction filtering algorithm and for generating a second resultant based on the first resultant using a nonlinear post-filtering algorithm. An audio output signal is then generated from the second resultant using a feature recovery algorithm.The audio output signal is transmitted via the communication link to a loudspeaker, which may be located at a remote location.

[0009] The feature recovery algorithm includes a module based on a deep neural network (DNN-based module) that contains: an STFT (short-term Fourier transform); several convolutional layers; a first LSTM (long-short-term memory layer); a second LSTM layer; a dense layer; several transposed convolutional layers; and an ISTFT (inverse short-term Fourier transform layer).

[0010] Another aspect of the invention is that the STFT transforms the audio input signal from an amplitude range into a frequency range.

[0011] Another aspect of the invention includes the fact that the STFT transforms the audio input signal into the frequency domain as a 2-channel sequence with a real part and an imaginary part.

[0012] Another aspect of the invention includes the multiple convolutional layers comprising a first convolutional layer with a 2-channel input with 256 features and a 32-channel output with 128 features; a second convolutional layer with a 32-channel input with 128 features and a 64-channel output with 64 features; a third convolutional layer with a 64-channel input with 64 features and a 128-channel output with 32 features; a fourth convolutional layer with a 128-channel input with 32 features and a 128-channel output with 16 features; and a fifth convolutional layer with a 128-channel input with 16 features and a 256-channel output with 8 features. and a sixth convolution layer with a 256-channel input with 8 features and a 256-channel output with 4 features.

[0013] Another aspect of the invention includes the fact that the 256-channel output with 4 features, which is output by the sixth convolution layer, is provided as an input to the first LSTM layer.

[0014] Another aspect of the invention includes the fact that each of the multiple folding layers has a core of size (2, 9) and a step size of size (1, 2).

[0015] Another aspect of the invention includes the provision of an input of the first convolution layer as an input to the ISTFT.

[0016] Another aspect of the invention includes the fact that the output of the sixth convolution layer is provided as input to the first LSTM layer.

[0017] Another aspect of the invention includes the fact that the first LSTM layer has 256 states.

[0018] Another aspect of the invention includes the fact that the second LSTM layer has 256 states.

[0019] Another aspect of the invention includes the fact that the output of the second LSTM layer is provided as input to a dense layer.

[0020] Another aspect of the invention includes the fact that the multiple transposed convolutional layers comprise a sixth transposed convolutional layer with a 512-channel input with 4 features and a 256-channel output with 8 features; a fifth transposed convolutional layer with a 512-channel input with 8 features and a 128-channel output with 16 features; a fourth transposed convolutional layer with a 256-channel input with 16 features and a 128-channel output with 32 features; a third transposed convolutional layer with a 256-channel input with 32 features and a 64-channel output with 64 features; and a second transposed convolutional layer with a 128-channel input with 64 features and a 32-channel output with 128 features. and feature a first transposed convolution layer with a 64-channel input with 128 features and a 2-channel output with 256 features.

[0021] Another aspect of the invention includes the fact that the output of the dense layer is provided as input to the sixth transposed folding layer.

[0022] Another aspect of the invention includes the fact that each of the several transposed folding layers has a core of size (2, 9) and a step size of size (1, 2).

[0023] Another aspect of the invention includes providing the output of the first transposed convolution layer as an input to the ISTFT to effect feature recovery.

[0024] Another aspect of the invention includes the fact that the output of the first convolution layer is provided as an input to the first transposed convolution layer.

[0025] Another aspect of the invention includes the fact that the output of the second convolution layer is provided as an input to the second transposed convolution layer.

[0026] Another aspect of the invention includes the fact that the output of the third convolution layer is provided as an input to the third transposed convolution layer.

[0027] Another aspect of the invention includes the fact that the output of the fourth convolution layer is provided as an input to the fourth transposed convolution layer.

[0028] Another aspect of the invention includes the fact that the output of the fifth convolution layer is provided as an input to the fifth transposed convolution layer.

[0029] Another aspect of the invention includes the fact that the output of the sixth convolution layer is provided as an input to the sixth transposed convolution layer.

[0030] Another aspect of the invention includes the fact that the ISTFT transforms the transposed audio input signal together with the output of the first transposed convolution layer from a frequency domain into an amplitude domain to generate the audio output signal.

[0031] The inventive method for processing an audio input signal comprises: capturing an audio input signal via a microphone; subjecting the audio input signal to a linear noise reduction filtering algorithm to generate a first resultant; subjecting the first resultant to a nonlinear post-filtering algorithm to generate a second resultant; generating an audio output signal by subjecting the second resultant to a feature restoration algorithm; and controlling a loudspeaker in response to the audio output signal.

[0032] Another aspect of the invention comprises a system for processing speech input, including a microphone, a controller, and a loudspeaker, wherein the microphone is configured to capture a speech input signal and transmit the speech input signal to the controller; and wherein the controller is functionally connected to the loudspeaker. The controller includes executable code to subject the speech input signal to a linear noise reduction filtering algorithm to generate a first resultant; to subject the first resultant to a nonlinear post-filtering algorithm to generate a second resultant; to generate an audio signal by subjecting the second resultant to a feature restoration algorithm; and to control the loudspeaker in response to the speech output signal.

[0033] The above summary is not intended to represent every possible embodiment or aspect of the present invention. Rather, the preceding summary is intended to illustrate by way of example some of the novel aspects and features disclosed herein. The above features and advantages, and further features and advantages of the present invention, will readily become apparent from the following detailed description of representative embodiments and embodiments of the present invention when taken together with the accompanying drawings and the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] One or more embodiments will now be described by way of example with reference to the attached drawings; they show: Fig. 1 schematically a microphone, a controller and a communication link which can be coupled with a remote loudspeaker, according to the invention; Fig. 2 schematic elements of a noise reduction routine for processing an audio input signal according to the invention. Fig. 3 schematic elements of a feature recovery algorithm which includes a module of a deep neural network (DNN module) for processing an audio input signal as part of a noise reduction routine, according to the invention. Fig. 4 schematic elements relating to a training module for training a module of a deep neural network (DNN module) for processing an audio input signal according to the invention.

[0035] The accompanying drawings are not necessarily to scale and may represent a somewhat simplified depiction of various preferred elements of the present invention as disclosed herein, including, for example, specific dimensions, orientations, locations, and shapes. Details associated with such elements are partly determined by the specific intended application and usage environment. DETAILED DESCRIPTION

[0036] Now, based on the drawings provided to illustrate certain exemplary embodiments, it is stated that Fig. Figure 1 schematically depicts a system 100 comprising a microphone 20 and a controller 10 capable of communicating with a remote audio loudspeaker 70 via a communication link 60. According to one embodiment, the remote audio loudspeaker 70 is located outside the system 100. The system 100 includes a noise reduction routine 200 to manage an audio input signal 15, thereby reducing audible background noise levels and improving speech intelligibility. The term "speech intelligibility" refers to speech clarity, i.e., the degree to which speech sounds can be correctly identified and understood by a listener.

[0037] The microphone 20 can be any device containing a transducer capable of converting audible sound into an electrical signal in the form of an audio input signal 15. The communication link 60 can be a direct wired point-to-point connection, a networked communication bus connection, a wired connection, or another type of communication link.

[0038] The controller 10 includes a receiver 30, a processor 40 and a memory 50, wherein the memory 50 contains an embodiment of the noise reduction routine 200 and provides data storage.

[0039] The term "controller" and related terms such as microcontroller, control unit, processor, etc., refer to one or more combinations of one or more application-specific integrated circuits (ASICs), free programmable logic gates (FPGAs), electronic circuits, central processing units, e.g., microprocessors, and associated transitive and non-transient memory components in the form of storage and data storage devices (read-only, programmable read-only, write-read, hard disk drive, etc.).The non-transitory memory component is capable of storing machine-readable instructions in the form of one or more software or firmware programs or software or firmware routines, one or more combination logic circuits, one or more input / output circuits and input / output devices, a signal conditioning circuit, a buffer circuit arrangement, and other components that can be accessed by one or more processors and that are executed by a. or multiple processors can be executed to provide and store a described functionality. The one or more input / output circuits and input / output devices contain analog-to-digital converters and associated devices that monitor inputs from sensors, such inputs being monitored at a preset sampling rate or in response to a trigger event. Software, firmware, programs, instructions, control routines, code, algorithms, and similar terms mean sets of instructions, including calibrations and lookup tables, that can be executed by a controller. Each controller executes one or more control routines to provide desired functions. Routines can be executed at regular intervals, for example, every 100 microseconds during continuous operation. Alternatively, routines can be executed in response to the occurrence of a trigger event.Communication between controllers, actuators, and / or sensors and the remote audio speaker 70 can be achieved using a direct-wired point-to-point connection, a networked communication bus connection, a wireless connection, or another communication link. This communication involves the exchange of data signals, including, for example, electrical signals via a conductive medium; electromagnetic signals through the air; optical signals via fiber optics; etc. The data signals can include discrete analog and / or digitized analog signals representing sensor inputs, actuator commands, and communication between controllers.

[0040] The term "signal" refers to a physically distinguishable indicator that transmits information and can be a suitable signal form (e.g., electrical, optical, magnetic, mechanical, or electromagnetic) such as DC, AC, sine wave, triangular wave, square wave, vibration, and the like, which is capable of passing through a medium.

[0041] Fig. Figure 2 schematically provides elements of the noise reduction routine 200 for processing the audio input signal 15, which includes a linear noise reduction algorithm 210, a nonlinear post-filter algorithm 240 and a feature recovery algorithm 300.

[0042] The linear noise reduction algorithm 210 incorporates acoustic echo suppression (AEC) 220 and beamforming (BF) 230. AEC 220 is a digital signal processing technique for identifying and suppressing acoustic echoes, which are reduced to function as an algorithm. BF 230 is a digital signal processing technique that uses spatial information to reduce ambient noise power, thereby improving the signal-to-noise ratio. In one embodiment, AEC 220 precedes BF 230. Alternatively, BF 230 can precede AEC 220. Acoustic echo suppression and beamforming are acoustic signal processing techniques familiar to experienced practitioners.

[0043] The linear noise reduction algorithm 210 generated a first resulting signal 235, which is provided as input to the nonlinear post-filtering algorithm (NLP algorithm) 240. The NLP algorithm 240 improves the noise reduction level by using nonlinear filtering to reduce residual noise and echoes. NLP is an acoustic signal processing technique familiar to experienced practitioners.

[0044] The NLP algorithm 240 generates a second resulting signal 245, which is provided as input to the feature recovery algorithm 300. Based on this second resulting signal 245, the feature recovery algorithm 300 generates the audio output signal 55. The DNN-based feature recovery algorithm 300 is placed after the post-filtering module to simplify tuning and improve speech quality.

[0045] Fig. Figure 3 schematically represents elements of the feature recovery algorithm 300 for processing the audio input signal 15 as part of the noise reduction routine 200. The feature recovery algorithm 300 is composed of a module of a deep neural network (DNN module) containing a short-term Fourier transform layer (STFT layer) 310, several convolutional layers 320, a first long-short-term memory layer (LSTM layer) 330, a second LSTM layer 332, a dense layer 340, several transposed convolutional layers 350, and an ISTFT layer 370.

[0046] The STFT and ISTFT layers 310 and 370 are each a sequence of Fourier transforms of a windowed signal, providing time-localized frequency information for situations where the frequency components of a signal vary over time. A recurrent neural network (RNN) is a time-series version of an artificial neural network (ANN) designed to process sequences of data, such as sound. In speech processing, an RNN-based deep neural network (DNN) utilizes strong correlations between speech time and speech frequency for noise reduction and blind source separation. This capability can be leveraged for the recovery problem, leading to simplified tuning of the post-filter module at lower ambient noise levels to achieve improved speech quality in the form of speech intelligibility.

[0047] The first and second long-short-term memory (LSTM) layers 330 and 332 are a type of recurrent neural network commonly used for tasks such as text-to-speech or natural language processing. They possess a recurrent state that is updated each time new data is fed through the network. In this way, the LSTM layers have a memory function.

[0048] The STFT layer 310 transforms the audio input signal 15 in the form of a 2-channel sequence with a real part and an imaginary part from an amplitude domain into a frequency domain.

[0049] According to one embodiment, the multiple convolutional layers 320 include a first convolutional layer 321 with a 2-channel input with 256 features and a 32-channel output with 128 features; a second convolutional layer 322 with a 32-channel input with 128 features and a 64-channel output with 64 features; a third convolutional layer 323 with a 64-channel input with 64 features and a 128-channel output with 32 features; a fourth convolutional layer 324 with a 128-channel input with 32 features and a 128-channel output with 16 features; a fifth convolutional layer 325 with a 128-channel input with 16 features and a 256-channel output with 8 features; and a sixth convolution layer 326 with a 256-channel input with 8 features and with a 256-channel output with 4 features.

[0050] According to one embodiment, each of the multiple convolution layers 320 has a kernel of size (2, 9) and a step size of size (1, 2). The kernel is a filter used to extract features from the data and is a matrix that traverses the input data, performs a scalar product on a subset of the input data, and has an output as the matrix of scalar products. The step size controls how the filter convolves around the input volume.

[0051] The 256-channel output with 4 features (327) output by the sixth convolution layer 326 is provided as an input to the first LSTM layer 330, which has 256 states.

[0052] An input from the first convolution layer 321 is provided as an input to the ISTFT layer 370.

[0053] An output from the first LSTM layer 330 is provided as input to the second LSTM layer 332, and an output from the second LSTM layer 332 is provided as input to the dense layer 340.

[0054] An output of the dense layer 340 is provided as input (357) to the multiple transposed convolution layers 350, specifically to a sixth convolution layer 326.

[0055] The multiple transposed convolutional layers 350 contain a sixth transposed convolutional layer 356 with a 512-channel input with 4 features and a 256-channel output with 8 features; a fifth transposed convolutional layer 355 with a 512-channel input with 8 features and a 128-channel output with 16 features; a fourth transposed convolutional layer 354 with a 256-channel input with 16 features and a 128-channel output with 32 features; a third transposed convolutional layer 353 with a 256-channel input with 32 features and a 64-channel output with 64 features; a second transposed convolutional layer 352 with a 128-channel input with 64 features and a 32-channel output with 128 features; and a first transposed convolution layer 351 with a 64-channel input with 128 features and with a 2-channel output with 256 features.

[0056] According to one embodiment, each of the several transposed folding layers 350 has a core of size (2, 9) and a step size of size (1, 2).

[0057] An output of the first convolution layer 321 is provided as an input to the first transposed convolution layer 351.

[0058] An output of the second convolution layer 322 is provided as an input to the second transposed convolution layer 352.

[0059] An output of the third convolution layer 323 is provided as an input to the third transposed convolution layer 353.

[0060] An output of the fourth convolution layer 324 is provided as an input to the fourth transposed convolution layer 354.

[0061] An output of the fifth convolution layer 325 is provided as an input to the fifth transposed convolution layer 355.

[0062] An output of the sixth convolution layer 326 is provided as an input to the sixth transposed convolution layer 356.

[0063] The output of the first transposed convolution layer 251 is added to the input of the first convolution layer 321 and the sum is provided as an input to the ISTFT layer 370 to effect feature recovery when generating the audio output signal 55.

[0064] It will be acknowledged that the number of convolution layers 320, the number of features and channels assigned to each convolution layer 320, the number of transposed convolution layers 350, the number of features and channels assigned to each transposed convolution layer 350, the kernel sizes and step sizes, the number, type and size of the RNN layers (330, 332) and the number and size of the dense layer (340) are application-specific and are chosen on the basis of factors relating to computing speed, processor capabilities, sound quality, etc.

[0065] Fig. Figure 4 schematically presents elements relating to a training module 400 for training an embodiment of the module of a deep neural network (DNN module) based on Fig. The feature recovery algorithm 300 described in Section 3 processes an audio input signal 15. Inputs to the training module 400 include an audio input signal in the form of clean speech 411 and an audio input signal in the form of noise 412, e.g., white noise, road noise, babble, etc., both provided in an amplitude range. The clean speech 411 and the noise 412 are input to an STFT layer 410, which transforms them into the frequency domain as transformed clean speech 411' and transformed noise 412'.

[0066] The transformed clean speech 411' and the transformed noise 412' are added to form noisy speech 415. The noisy speech 415 and the transformed noise 412' are inputted to the NLP 420, which improves the noise reduction level by using nonlinear filtering to attenuate the noise level. Outputs of the NLP 420 include residual noise 422 and a combination of distorted speech and residual noise 424. The residual noise 422 is added to the transposed clean speech 411' to form an initial input 426. The initial input 426, in the form of residual noise 422 added to the transformed clean speech 411', and the combination of distorted speech and residual noise 424, are used as inputs to the NLP 420. Fig. The feature recovery algorithm 300 described in section 3 was provided to effect training.

[0067] This arrangement of inputs to training module 400 causes the feature recovery algorithm 300 to be trained to recover the missing speech features without affecting the noise levels. The residual noise signal is generated by processing the noise signal in accordance with noisy speech processing. The deep learning approach described here unifies the feature extraction process across multiple layers of the neural network. During the training process, the parameters of the neural network are learned, and then real-time sound is fed into the trained neural network in real time to achieve speech feature recovery.

[0068] The concepts described here create a system that uses a speech feature restoration module instead of a flawlessly tuned primary frequency (PF). The feature restoration module monitors the restoration of the original speech quality, enabling both improved noise reduction and voice quality that cannot be achieved using other known methods. In the case of flawless restoration, the PF can be configured to output the desired noise level independently of the added desired speech distortion.

[0069] Embodiments of the present invention can be embodied as a device, a method, or a computer program product. Accordingly, the present invention can take the form of a complete hardware embodiment, a complete software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, all of which may herein be generally referred to as a "module" or "system." Furthermore, the present invention can take the form of a computer program product embodied in a specific medium with an expression comprising computer-usable program code embodied in that medium.

[0070] The flowchart and block diagrams in the flowcharts represent the architecture, functionality, and operation of possible implementations of systems, processes, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams can represent a module, section, or portion of code comprising one or more executable instructions for implementing the one or more specified logical functions. It is further noted that each block in the block diagrams and / or flowchart representations, and combinations of blocks in the block diagrams and / or flowchart representations, can be implemented by hardware-based systems with dedicated functions that perform the specified functions or activities, or by combinations of hardware with dedicated functions and computer instructions.These computer program instructions may also be stored in a computer-readable medium that can instruct a computer or other programmable data processing device to function in a particular way so that the instructions stored in the computer-readable medium produce a manufactured article containing a set of instructions that implements the function / activity specified in the flowchart and / or in the one or more block diagram blocks.

Claims

[1] System (100) for processing an audio input signal (15), wherein the system (100) comprises: a microphone (20), a controller (10), a data storage device (50) and a communication link (60) with a remote audio speaker (70); wherein the microphone (20) is configured to capture and generate the audio input signal (15) and to transmit the audio input signal (15) to the controller (10); wherein the controller (10) is functionally connected to the communication link (60); and wherein the data storage (50) contains instructions that can be executed by the controller (10), wherein the instructions contain: Generating a first resultant (235) based on the audio input signal (15) via a linear noise reduction filtering algorithm; Generating a second resultant (245) based on the first resultant (235) using a nonlinear post-filtering algorithm; Generating an audio output signal based on the second resultant (245) via a feature recovery algorithm (300); and Transmitting the audio output signal to the remote audio speaker (70) via the communication link (60); characterized by , that the feature recovery algorithm (300) comprises a module based on a deep neural network (DNN-based module) which includes: a STFT (short-term Fourier transform) (310); several convolutional layers (320); a first LSTM (long-short-term memory layer) (330); a second LSTM (332); a dense layer (340); several transposed convolutional layers (350); and an inverse STFT (ISTFT) (370). [2] System (100) according to claim 1, wherein the STFT (310) transforms the audio input signal (15) from an amplitude range into a frequency range. [3] System (100) according to claim 2, wherein the STFT (310) transforms the audio input signal (15) into the frequency domain using a 2-channel sequence with a real part and an imaginary part. [4] System (100) according to claim 1, wherein the multiple folding layers (320) comprise: a first convolution layer (321) with a 2-channel input with 256 features and with a 32-channel output with 128 features; a second convolution layer (322) with a 32-channel input with 128 features and with a 64-channel output with 64 features; a third convolution layer (323) with a 64-channel input with 64 features and with a 128-channel output with 32 features; a fourth convolution layer (324) with a 128-channel input with 32 features and with a 128-channel output with 16 features; a fifth convolution layer (325) with a 128-channel input with 16 features and with a 256-channel output with 8 features; and a sixth convolution layer (326) with a 256-channel input with 8 features and with a 256-channel output with 4 features. [5] System (100) according to claim 4, wherein the 256-channel output with 4 features, which is output by the sixth convolution layer (326), is provided as an input to the first LSTM layer (330). [6] System (100) according to claim 4, wherein each of the multiple folding layers (320) has a core of size (2, 9) and a step size of size (1, 2). [7] System (100) according to claim 4, wherein an output of the first convolution layer (321) is provided as an input to the ISTFT (370). [8] System (100) according to claim 4, wherein the output of the sixth convolution layer (326) is provided as input to the first LSTM layer (330). [9] Method for processing an audio input signal (15) comprising: that an audio input signal (15) is captured via a microphone (20); that the audio input signal (15) is subjected to a linear noise reduction filtering algorithm to generate a first resultant (235); that the first resultant (235) is subjected to a nonlinear post-filtering algorithm to generate a second resultant (245); that by subjecting the second resultant (245) to a feature recovery algorithm (300), an audio output signal is generated; and that a loudspeaker is controlled in response to the audio output signal; characterized by , that the feature recovery algorithm (300) comprises a module based on a deep neural network (DNN-based module) which includes: a STFT (short-term Fourier transform) (310); several convolutional layers (320); a first LSTM (long-short-term memory layer) (330); a second LSTM (332); a dense layer (340); several transposed convolutional layers (350); and an inverse STFT (ISTFT) (370).

Images