Systems and methods for interference signal reduction

Abstract

An interference signal reduction system and a method of interference signal reduction are described. The input signal includes a plurality of signal segments (frames). The input signal is provided to a machine learning model that is trained to output an estimate of a desired target signal from the input signal. Each signal segment may be classified as only a target signal segment, only an interfering signal segment, or an undefined signal segment. The machine learning model may be adjusted based on the target only signal segment and the interfering only signal segment.

Description

Technical Field

[0001] The present disclosure relates to an interference signal reduction system and an interference signal reduction method. Background Art

[0002] A model-based interference signal reduction method has been applied to train a machine learning model to estimate a desired target signal based on features extracted from observed sensor signals. Machine learning model training is performed offline using representative data that approximates actual sensor data. Once trained, the machine learning model is deployed on a processing device where it performs interference signal reduction. Summary of the invention

[0003] Various aspects of the disclosure are defined in the accompanying claims. In a first aspect, a method of interference signal reduction is provided, the method comprising: receiving an input signal comprising a plurality of signal segments; providing the input signal to a machine learning model, the machine learning model being configured to output an estimate of a target signal based on the input signal; wherein the method further comprises: classifying each input signal segment of the plurality of signal segments into one of a target signal-only segment, an interference signal-only segment, and an optional undefined signal segment; and adjusting the machine learning model based on the target signal-only segment and the interference signal-only segment.

[0004] In some embodiments, classifying each input signal segment further includes: estimating a target signal-to-interference ratio of the input signal segment; classifying the input signal segment as a target signal-only segment in response to the target signal-to-interference ratio exceeding a first threshold; and classifying the input signal segment as an interference signal-only segment in response to the target signal-to-interference ratio being less than a second threshold that is lower than the first threshold.

[0005] In some embodiments, adjusting the machine learning model further comprises:

[0006] A plurality of target plus interference signal segments are generated by mixing a selected target-only signal segment with a selected interference-only signal segment at a target signal-to-interference ratio to obtain each target plus interference signal segment. In some embodiments, the input signal further comprises at least one of an audio signal and a video signal. In some embodiments, the input signal further comprises an audio signal and a video signal, and wherein classifying the input signal further comprises: determining a noise power according to at least one audio component of the input signal segment; classifying the input signal segment as a target-only signal segment in response to the noise power being less than a noise power threshold and voice activity being detected from a video component of the input signal segment; and classifying the input signal segment as an interference-only signal segment in response to the noise power being greater than the noise power threshold and voice activity not being detected from the video component of the input signal segment. In some embodiments, adjusting the machine learning model further comprises: generating a plurality of target plus interference signal segments; wherein the audio component of the target plus interference signal segment is generated by mixing an audio component of a selected target-only signal segment with an audio component of a selected interference-only signal segment at a target signal-to-interference ratio; and the video component of the target plus interference signal segment comprises the video component of the selected target-only signal segment.

[0007] In some embodiments, adjusting the machine learning model further includes: constructing a target plus interference data set from the multiple target plus interference signal segments; constructing a target signal-only data set from the corresponding selected target signal-only segments of each target plus interference signal segment; and applying the target plus interference data set and the target signal-only data set to the machine learning model.

[0008] In some embodiments, the input signal comprises a digital audio signal, wherein the input signal segment comprises an audio packet, the audio packet comprises a plurality of audio samples, and wherein classifying the input signal segment further comprises: classifying the input signal segment as a target signal-only segment in response to correctly receiving the audio packet; and classifying the input signal segment as an interference signal-only segment in response to incorrectly receiving the audio packet.

[0009] In some embodiments, the method further comprises: updating at least one parameter of a packet loss pattern model with the interference-only signal segments. In some embodiments, adjusting the machine learning model further comprises: constructing a target signal-only dataset from a plurality of selected target signal-only segments; constructing a target plus interference dataset by muting a subset of the plurality of selected target signal-only segments according to the pattern loss packet model; and applying the target plus interference dataset and the target signal-only dataset to the machine learning model.

[0010] In some embodiments, the method further comprises: storing only target signal segments and only interference signal segments in separate databases. In some embodiments, after storing the additional signal segments, modifying the database to remain within a predefined size. In some embodiments, the method further comprises: receiving another input signal comprising a plurality of additional signal segments; providing the other input signal to the machine learning model, the machine learning model further configured to output an estimate of another target signal based on the other input signal; classifying each of the plurality of additional signal segments as one of a target signal segment only, an interference signal segment only, and an optional undefined signal segment; and adjusting the machine learning model based on the target signal segment only and the interference signal segment only.

[0011] In a second aspect, an interference signal reduction system is provided, comprising: a signal input configured to receive an input signal comprising a plurality of signal segments; an interference signal reduction system output; a target signal estimator configured to output an estimate of a target signal based on the input signal, the target signal estimator comprising a machine learning model and having a target signal estimator input coupled to the signal input, a target signal estimator output coupled to the interference signal reduction system output, and an adapter input; a signal classifier having a signal classifier input coupled to the signal input, a target signal output, and an interference signal output; an adapter having a target signal input coupled to the target signal output, an interference signal input coupled to the interference signal output, and a model adapter output coupled to the adapter input; wherein the signal classifier is configured to classify each input signal segment of the plurality of signal segments into one of a target signal segment only, an interference signal segment only, and an optional undefined signal segment; and the adapter is configured to adjust the machine learning model based on the target signal segment only and the interference signal segment only.

[0012] In some embodiments, the signal classifier is further configured to: estimate a target signal-to-interference ratio of the input signal segment; classify the input signal segment as a target signal-only segment in response to the target signal-to-interference ratio exceeding a first threshold; and classify the input signal segment as an interference signal-only segment in response to the target signal-to-interference ratio being less than a second threshold that is lower than the first threshold.

[0013] In some embodiments, the adapter is further configured to generate a plurality of target-plus-interference signal segments by mixing selected target-only signal segments with selected interference-only signal segments at a target signal-to-interference ratio to obtain each target-plus-interference signal segment.

[0014] In some embodiments, the input signal further includes an audio signal and a video signal, and wherein the signal classifier is further configured to: determine a noise power based on at least one audio component of the input signal segment; classify the input signal segment as a target signal-only segment in response to the noise power being less than a noise power threshold and voice activity being detected from the video component of the input signal segment; and classify the input signal segment as an interference signal-only segment in response to the estimated noise power being greater than the noise power threshold and no voice activity being detected from the video component of the input signal segment.

[0015] In some embodiments, the adapter is further configured to generate multiple target plus interference signal segments; wherein the audio component of the target plus interference signal segment is generated by mixing the audio component of a selected target-only signal segment with the audio component of a selected interference-only signal segment at a target signal-to-interference ratio; and the video component of the target plus interference signal segment includes the video component of the selected target-only signal segment.

[0016] In some embodiments, the adapter is further configured to: construct a target plus interference data set from the multiple target plus interference signal segments; construct a target signal-only data set from the corresponding selected target signal-only segments of each target plus interference signal segment; and apply the target plus interference data set and the target signal-only data set to the machine learning model.

[0017] In some embodiments, the input signal comprises a digital audio signal, the input signal segment comprises an audio packet, the audio packet comprises a plurality of audio samples, and wherein the signal classifier is further configured to: classify the input signal segment as a target signal-only segment in response to correctly receiving the audio packet; classify the input signal segment as an interference signal-only segment in response to incorrectly receiving the audio packet; and update at least one parameter of a packet loss pattern model with the interference signal-only segment; and wherein the adapter is further configured to: construct a target signal-only dataset from a plurality of selected target signal-only segments; construct a target plus interference dataset by muting a subset of the plurality of selected target signal-only segments according to a pattern loss packet model; and apply the target plus interference dataset and the target signal-only dataset to the machine learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In the drawings and description, like reference numerals refer to like features. Embodiments will now be described in detail only by way of example as illustrated by the accompanying drawings, in which:

[0019] Figure 1 An interfering signal reduction system according to an embodiment is shown.

[0020] Figure 2A method for reducing interference signals according to an embodiment is shown.

[0021] Figure 3 An interfering signal reduction system according to an embodiment is shown.

[0022] Figure 4 A method of signal classification for interference signal reduction according to an embodiment is shown.

[0023] Figure 5 A method of adjusting an interference reduction system according to an embodiment is shown.

[0024] Figure 6 A method of audio and video signal classification for interference signal reduction according to an embodiment is shown.

[0025] Figure 7 A method of adjusting an interference reduction system according to an embodiment is shown.

[0026] Figure 8 A method for digital audio signal classification with interference signal reduction according to an embodiment is shown.

[0027] Fig. 9 A method of adjusting an interference reduction system according to an embodiment is shown.

[0028] It should be noted that the figures are diagrammatic and not drawn to scale. For clarity and convenience in the drawings, the relative sizes and proportions of the various parts of these figures have been shown as enlarged or reduced in size. The same reference numerals are generally used to refer to corresponding or similar features in modified and different embodiments. DETAILED DESCRIPTION

[0029] Figure 1 An interference signal reduction system 100 is shown according to an embodiment. A signal input 110 may be connected to a signal classifier input of a signal classifier 112 and a target signal estimator input of a target signal estimator 122, which is implemented as a machine learning model. The signal classifier 112 may have a target signal output 114 and an interference signal output 116 connected to respective target signal inputs and interference signal inputs of an adapter 118. The adapter 118 may have an adapter output 120 connected to an adapter input of the target signal estimator 122. The target signal estimator 122 has a target signal estimator output connected to an interference signal reduction system output 124.

[0030] In operation, a signal may be received at signal input 110. The signal may include a target signal from source 102, which may be degraded by one or more interfering sources 104, as shown by interference process 108.

[0031] Figure 2 A method 150 of interference signal reduction is shown, which may be implemented, for example, by the interference reduction system 100. In step 152, an input signal including a target signal from a source 102, which may be degraded by one or more interference sources, may be received at the signal input 110. The received signal may initially be detected by a sensor (not shown), which may be one or more of a microphone, a camera, and an antenna, depending on the type of target signal.

[0032] The received signal may be processed, for example, by converting an analog signal into a digital signal and provided to a target signal estimator 122, which is implemented as a machine learning model configured to output an estimate of the target signal based on the input signal (step 154). The input signal may be considered to be composed of a plurality of signal segments. For example, in the digital domain, an input signal segment may be composed of a number N of signal samples corresponding to a predetermined duration. The target signal estimator 122 may then output an estimate of the target (desired) signal based on the received signal at the interference signal reduction system output 124 for each signal segment.

[0033] In step 156, the signal classifier 112 may classify the signal segments of the input signal segments into one of only target signal segments, only interference signal segments, and optionally undefined signal segments. Only the target signal segments may be output on the target signal output 114, and only the interference signal segments may be output on the interference signal output 116. Undefined signal segments may be discarded. The adapter 118 may adjust the machine learning model based on only the target signal segments and only the interference signal segments by constructing a training data set for updating the parameters of the machine learning model. The update process typically requires one or more iterations, which consists of a forward step of applying the input data to the machine learning model and observing the output, and a backpropagation step, in which the machine learning model outputs a reference output from the training set, and adjusts the model parameters to minimize the difference between the reference output and the model output. In some examples, the forward-backward iteration may be applied to a copy of the machine learning model in the adapter 118, which then replaces the machine learning model in the target signal estimator 122. In other examples, the forward-backward iteration may be applied directly to the machine learning model in the target signal estimator 122 by the adapter 118 during the downtime period. Steps 152 and 154 may be implemented in software running as a foreground process on a processor. Steps 156 and 158 may be implemented in software running as a background process on a processor.

[0034] The interference reduction system 100 and method 150 may allow on-system machine learning for interference signal reduction. In some examples, the system 100 may be implemented as a system on a chip (SoC), a microprocessor, a microcontroller, or other integrated circuit device. The machine learning model may be trained or adjusted online using actual sensor signal data collected on the system. The interference reduction system 100 and method 150 may allow the interference signal reduction machine learning model to be personalized for potential dynamic targets and interference signal characteristics specific to the device / system, user, and / or environment, i.e., personalized machine learning.

[0035] For the example of implementing system 100 on a device, system 100 allows learning or tuning of an interference signal reduction machine learning model to be performed entirely on the device without requiring data exchange or processing outside the device, i.e., privacy-preserving machine learning.

[0036] Figure 3 An interference signal reduction system 200 according to an embodiment is shown. Signal inputs 222, 224, 226 for different signal types may be connected to corresponding signal classifier inputs of a signal classifier 228 and corresponding target signal estimator inputs of a target signal estimator 248, the target signal estimator comprising a feature extractor 230 and an interference reduction machine learning model 254. The feature extractor 230 transforms the signal input into input features output on a feature extraction output 232, which are inputs to the interference reduction ML model 254. The feature extractor 230 may calculate one or more feature vectors using time or spectral processing and transformation. Several feature vectors may be combined or concatenated to form a single input feature vector. The signal classifier 228 may have a target signal output 234 and an interference signal output 238, the target signal output being connected to the input of a target signal storage device 240 (memory) for a database of only target signal segments, and the interference signal output being connected to the input of an interference signal storage device 242 (memory) for a database of only interference signal segments. In some examples, the target signal storage and the interference signal storage may be included in the model adapter 250. The signal classifier 228 may have an undefined signal output 236. An output 244 of the target signal storage 240 may be connected to a target signal input of the model adapter 250. An output 246 of the target signal storage 242 may be connected to a target signal input of the model adapter 250. The model adapter 250 may have a model adapter output 252 connected to an adapter input of an interference reduction machine learning model 254. The feature extractor 230 has a feature extraction output 232 connected to an input of the interference reduction machine learning model 254. An output of the interference reduction machine learning model 254 is connected to the interference signal reduction system output 206.

[0037] The target source 202 and the interference source 204 combine during the physical interference process 208, thereby generating signals 210, 212, 214, which can be detected by corresponding sensors 216, 218, 220 that observe the desired target signal and the undesired interference signal. The goal of the interference signal reduction system 200 is to remove the undesired interference signal while retaining the desired target signal based on the observed sensor signal. One or more input signals can be received at the signal input 222, 224, 226. The received signal can initially be detected by a sensor, which depends on the type of target signal, such as a microphone 216, a camera 218, and an antenna 220. The interference signal reduction system 200 shows three different types of signals that can be processed, namely an audio signal 210, a video signal 212, and a radio frequency (RF) signal 214. In other examples, fewer or more signal types can be processed.

[0038] The received signal may be processed, for example, by converting the sensor signal into a digital signal and provided to a target signal estimator 248, which is configured to output an estimate of the target signal based on the input signal. Each of the input signal types may be considered to consist of a plurality of signal segments. For example, in the digital domain, an input signal segment may consist of a number N of signal samples corresponding to a predetermined duration. The target signal estimator 248 may then output an estimate of the target (desired) signal based on the received signal at the interference signal reduction system output 206 for each signal segment.

[0039] The input to the interference reduction machine learning model 254 consists of features extracted from the sensor signal by the feature extractor 230. In a background process, the interference reduction machine learning model 254 can be adjusted by applying a signal classification process by the signal classifier 228 and a model adjustment process by the model adapter 250. The signal classifier 228 can perform a relatively coarse classification based on the input sensor signal. The signal classifier 228 can detect the presence of only target signal or only interference signal conditions, and in these cases, the corresponding resulting sensor signal segments output on the target signal output 234 and the interference signal output 238 are stored as target signal segments or interference signal segments in the corresponding target signal storage device 240 or interference signal storage device 242, respectively. In all other signal conditions, the sensor signal segment is not stored because it does not represent a target signal or interference signal, but can be output on the undefined signal output 236. Because the signal classifier 228 is designed to detect relatively extreme signal conditions, no complex or adjustment methods are required; instead, a relatively simple and fixed implementation may be sufficient. During the model tuning process, the stored target signal representation and interference representation can be used to optimize the interference signal reduction performance of the machine learning model for these measured target signal and interference signal conditions.

[0040] Different embodiments may have selected implementations of the following aspects:

[0041] Sensor configuration: which sensors are used for the background signal classification process, and which sensors are used for the foreground interference reduction process.

[0042] Signal classification method: Which method is used to decide at which time instances the target signal representation or the interference signal representation will be collected.

[0043] Estimated signal representation: What type of signal representation is extracted from the on-device sensor signal.

[0044] Model adjustment method: Which method is used to update the machine learning model based on the collected target signal representation and / or interference signal representation.

[0045] Figure 4 A signal classification method 270 is shown, and Figure 5 A model adjustment method 300 is shown, which can be implemented, for example, by the interference reduction system 100, 200. For example, methods 270 and 300 can be used in the case where the target source is speech, where there are acoustic interference sources, such as ambient noise, or acoustic echo coupling from a speaker to a microphone. The sensor configuration is unimodal: the sensor consists of one or more microphones. In this case, all microphone signals will typically be used for both the signal classification background process and the interference reduction foreground process. The microphone sensor picks up a mixture of the desired speech signal and the undesired interference signal. Methods 270, 300 are for the case of a single microphone sensor. In other examples, similar methods can be applied in the case of multiple microphone sensors.

[0046] The signal classification process is based on the target signal-to-interference ratio estimation of the microphone sensor signal s(n), where n is a discrete time index. In a first step, the signal s(n) is buffered into N-sample signal frames, also referred to herein as signal segments, and the resulting signal segments s k =[s((k-1)*N),s((k-1)*N+1),…,s(k*N-1)], where k=1,2,…,K is the segment index. In step 272, the target signal-to-interference ratio can be estimated. The target signal-to-interference ratio estimation γ can be implemented using any state-of-the-art method, such as the statistical WADA-SNR algorithm k =g(s k ), the statistical WADA-SNR algorithm uses waveform amplitude distribution analysis to estimate the signal-to-noise ratio of the voice signal. The resulting estimate γ k This can then be used in simple threshold logic for signal classification:

[0047]

[0048] In step 274, the target signal-to-interference ratio is compared with the upper threshold T hi If the target signal-to-interference ratio exceeds the upper threshold, then the signal segment s k is classified as a target signal segment only (step 276), and the signal s k Otherwise, in step 278, the target signal-to-interference ratio is compared with the lower threshold T lo If the signal-to-interference ratio drops below the lower threshold T lo , then the signal segment s k The signal segment s is classified as an interference-only signal segment (step 280). k In other cases (step 282), the signal segment s k A signal classified as undefined may be a mixture of target and interference, or neither target nor interference is present.

[0049] Now turn to reference Figure 5 , the model adjustment method 300 uses the stored target signal representation and the interference signal representation in the synthetic signal mixing process to generate multiple target plus interference signal segments. The method starts at step 302. In step 304, the target signal and the interference signal are selected from the stored representation, and the two signals are mixed at a selected signal-to-interference ratio to generate a target plus interference signal segment. For example, the signal selection and the signal-to-interference ratio selection can be determined according to a predefined setting or in the form of pseudo-random number generation within a predefined range. The resulting target plus interference signal segment is stored, and an equivalent clean target signal is also stored as a reference. In step 306, a check is performed to determine whether the process has been repeated sufficiently D times. If it has been repeated sufficiently D times, then the target plus interference signal segment generation stops (step 308), otherwise the process returns to step 304 and repeats. Therefore, the synthetic signal mixing process is repeated D times, thereby generating a target plus interference mixed data set (Step 310) and the aligned clean target signal data set (Step 312).

[0050] In step 314, the target plus interference mixed data set M and the clean target signal data set C are used to update the machine learning model in an iterative supervised learning process. In each iteration of the supervised learning process, a batch of mixed signals M is constructed. i , which contains a subset of batch size B signals from the full dataset M. Thus, a batch of aligned clean target signals C is constructed i , which contains a subset of batch size B signals from the full dataset C. Given a model with parameters θ i-1The current machine learning model f() performs a model adjustment step with step size μ in each iteration i:

[0051] θ i =θ i-1 +μΔθ i ,

[0052] in

[0053]

[0054] And the data batch M i and C i Used to optimize the machine learning model parameter change Δθ i , in order to minimize the loss function Thus, the signal batch M is measured i and target batch C i is the interference cancellation quality on , and Δ is the optimization variable. This optimization problem can be solved, for example, using the known gradient descent method and back propagation technique.

[0055] The model tuning process may be considered complete according to stopping criteria, such as based on one or more of the following:

[0056] The maximum number of iterations I of the supervised learning process that have been run;

[0057] A target loss value or target loss value improvement has been achieved.

[0058] After the model tuning process is completed, the tuned machine learning model can replace the previous model in the foreground process, and interference reduction can be performed using the tuned machine learning model. The model tuning process can be scheduled to run at fixed time intervals, for example, using the latest sensor data, or whenever a sufficient amount of new target and interference data has been collected, or based on user input requesting an update to the model.

[0059] Figure 6 A signal classification method 350 is shown, and Figure 7A model adjustment method 400 is shown, which can be implemented, for example, by the interference signal reduction system 200. The target speech and acoustic interference sources are similar for methods 250, 300. However, in this case, the sensor configuration is multi-modal: the sensor consists of one or more microphones and, in addition, one or more camera sensors. All sensor signals are typically used for both signal classification background processes and interference reduction foreground processes. The microphone sensor picks up a mixture of the desired speech signal and the undesired acoustic interference signal. The camera sensor picks up relevant visual information related to the desired speech signal, such as the dynamic facial characteristics of the speaker, and is mostly insensitive to the acoustic interference signal. In this embodiment, an implementation of the present invention in the case of a single microphone sensor and a single camera sensor is described in detail. It should be understood that similar methods can be applied in the case of multiple microphones and camera sensors.

[0060] The signal classification process is now based on the microphone signal segments of the audio components also called input signal segments and the corresponding camera signal segment of the video component also referred to as the input signal segment The input signal segments are typically represented as RGB values ​​for different pixels in an image frame. One implementation of the signal classification process is to operate a binary voice activity detector on the camera signal segments. And the noise power estimate is operated on the microphone signal segment The voice activity detector can be implemented using a state-of-the-art detector based on head and lip movement tracking. Noise power estimation can be similarly implemented using a state-of-the-art noise power tracking algorithm. Signal classification is based on using a voice activity detector v k And the noise power estimate n k The combination logic of the two, for example:

[0061]

[0062] In step 352, in the input signal segment Video Components In step 354, based on the input signal segment At least one audio component Estimated noise power n k In step 356, a comparison is made between the noise power estimate and a threshold. noise And in the case where voice activity is detected in the subsequent step 358, the signal and The signal segment is classified as a target signal segment only (step 360) and can be stored in the target signal only database as a target signal representation. If no speech activity is detected in step 358, the signal segment is classified as an undefined signal segment (step 366). Returning to step 356, when the noise power estimate exceeds the threshold T noise And in the case where no voice activity is detected in the subsequent step 362, the signal and The signal is classified as an interference-only signal segment (step 364). Stored as interference signal; no signal stored Because it has no information related to acoustic interference.If speech activity is detected in step 362, the signal segment is classified as an undefined signal segment (step 366).

[0063] Now turn to reference Figure 7 , the model adjustment method 400 is most similar to that previously described for the unimodal speech enhancement case. The model adjustment method 400 uses the stored target signal representation and the interference signal representation in a synthetic signal mixing process to generate a plurality of target plus interference signal segments. The method begins at step 402. In step 404, a target signal and an interference signal are selected from the stored representations, and the audio components of the two signals are mixed at a selected signal-to-interference ratio to generate a target plus interference signal segment. The signal selection and signal-to-interference ratio selection may be determined, for example, according to a predefined setting or in the form of pseudo-random number generation within a predefined range. In step 406, the video component of the selected target-only signal segment is concatenated with the audio component of the target plus interference signal segment to form the target plus interference signal segment. The audio component of the target signal only is stored as a reference. In step 408, a check is performed to determine whether the process has been sufficiently repeated D times. If the process has been sufficiently repeated D times, then the target plus interference signal segment generation stops (step 410), otherwise the process returns to step 404 and repeats. Thus, the synthetic signal mixing process is repeated D times, thereby generating a target plus interference mixed data set M (step 412) and an aligned clean target signal data set C (step 414). The target plus interference mixed data set M (microphone and camera) and the clean target signal data set C (microphone) are used in step 416 to update the machine learning model in an iterative supervised learning process, similar to that previously described for method 300.

[0064] Figure 8 A signal classification method 450 is shown, and Fig. 9A model adjustment method 500 is shown, which can be implemented, for example, by the interference reduction system 100, 200. In this embodiment, the target signal is an audio stream, and there is an interference pattern of lost audio packets lost from the audio stream. Audio packet loss concealment is to restore these gaps in the audio stream caused by data transmission failures in a packet-switched network. The sensor configuration is unimodal: an audio packet stream is received by the interference reduction system. A foreground interference reduction machine learning model estimates the audio signal samples lost from the audio stream based on knowledge of the received signal samples.

[0065] refer to Figure 8 In step 452, a digital audio signal including one or more audio packets is received. In step 454, a check is performed to determine whether the audio packet is correctly received. If the audio packet is correctly received, then the target signal is present and no interfering signal is present. Signal frame s k =[s((k-1)*N),s((k-1)*N+1),…,s(k*N-1)] is classified as a target signal segment only (step 456). If the audio packet is not received correctly, then there is no target signal and an interference signal is present, and the interference signal segment is classified as an interference signal segment only (step 458). Information about the lost packet pattern is used to update the parameters of the packet loss pattern model, which serves as an interference signal representation (step 460). As an example, a two-state first-order Markov chain, also known as a Gilbert-Elliott channel, is often used to model packet loss patterns. This model has two parameters, namely, the transition probability p from the 'lost packet state' to the 'received packet state' and the transition probability q from the 'received packet state' to the 'lost packet state'. Updating the model parameters based on the observed packet loss pattern can be performed using state-of-the-art methods (e.g., maximum likelihood estimation). Method 450 can be repeated for multiple audio packets.

[0066] The model adjustment method 500 uses a generative data mixing process. The target signal-only segments are used to construct a clean target signal-only dataset C (step 502). A target plus interference dataset M is constructed by muting a plurality of selected subsets of the target signal-only segments according to a pattern loss packet model (step 504). Batches of pairs of clean audio streams from dataset C and corresponding audio streams from dataset M with model-generated packet loss can be used to update the interference reduction machine learning model using a supervised learning approach (step 506), as previously described for methods 300, 400.

[0067] Embodiments of systems and methods are described to train a model for interference signal reduction on a system using signal data collected by sensors. A relatively coarse signal classification stage can classify signal segments or frames into target-only and interference-only signal representations. These signal representations are used to optimize interference reduction model performance for these signal conditions through a model adjustment stage.

[0068] Some embodiments may be used for blind source separation, where there may be multiple target signal sources, all of which may be desired. The present invention may be applied in the context of blind source separation to detect the presence of different target sources using a signal classification process, and to adjust a blind source separation machine learning system based on an extracted representation of each of the target sources.

[0069] Embodiments may allow for model personalization for target and interference signal characteristics specific to the system / device, user, and / or environment.Where the interference reduction system is implemented on the device, the learning process is performed entirely on the device without data exchange or processing outside the device.

[0070] An interference signal reduction apparatus and a method for interference signal reduction are described. An input signal includes a plurality of signal segments (frames). The input signal is provided to a machine learning model, which is trained to output an estimate of a desired target signal based on the input signal. Each signal segment can be classified as a target signal segment only, an interference signal segment only, or an undefined signal segment. The machine learning model can be adjusted based on the target signal segment only and the interference signal segment only.

[0071] In some example embodiments, the instruction set / method steps described above are implemented as functions and software instructions embodied as an executable instruction set, which is implemented on a computer or a machine programmed and controlled with the executable instructions. Such instructions are loaded for execution on a processor (e.g., one or more CPUs). The term processor includes a microprocessor, a microcontroller, a processor module or subsystem (including one or more microprocessors or microcontrollers), or other control or computing devices. A processor may refer to a single component or multiple components.

[0072] In other examples, the instruction sets / methods shown herein and the data and instructions associated therewith are stored in corresponding storage devices, which are implemented as one or more non-transitory machine or computer readable or computer usable storage media. Such computer readable or computer usable storage media are considered to be part of an article (or product). An article or product may refer to any manufactured single component or multiple components. Non-transitory machine or computer usable media as defined herein do not include signals, but such media may be able to receive and process information from signals and / or other transient media.

[0073] Example embodiments of the materials discussed in this specification may be implemented in whole or in part via a network, computer or data-based device and / or service. These may include the cloud, the Internet, an intranet, a mobile device, a desktop computer, a processor, a lookup table, a microcontroller, a consumer device, an infrastructure, or other enabling devices and services. As may be used herein and in the claims, the following non-exclusive definitions are provided.

[0074] In one example, one or more instructions or steps discussed herein are automated. The term automation or automatically (and similar variations thereof) means the use of computers and / or mechanical / electrical devices to control the operation of a device, system, and / or process without human intervention, observation, effort, and / or decision making.

[0075] Although the attached claims are directed to specific feature combinations, it should be understood that the scope of the disclosure of the present invention also includes any novel feature or any novel feature combination or any generalized form thereof disclosed explicitly or implicitly herein, regardless of whether it relates to the same invention as currently claimed in any claim or whether it alleviates the same technical problem as any or all of the technical problems alleviated by the present invention.

[0076] Features that are described in the context of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features that are, for brevity, described in the context of a single embodiment may also be provided separately or in any suitable subcombination.

[0077] The applicants hereby give notice that new claims may be formulated to such features and / or combinations of such features during the prosecution of the present application or of any further application derived therefrom.

[0078] For the sake of completeness, it is also stipulated that the term "comprising" does not exclude other elements or steps, the term "a" does not exclude a plurality, a single processor or other unit may fulfil the functions of several components recited in the claims, and the reference signs in the claims should not be construed as limiting the scope of the claims.

Claims

1. A method for reducing interference signals, characterized in that: include: receiving an input signal comprising a plurality of signal segments; providing the input signal to a machine learning model, the machine learning model being configured to output an estimate of a target signal based on the input signal; wherein the method further comprises: classifying each input signal segment of the plurality of signal segments as one of a target signal segment only, an interfering signal segment only, and optionally an undefined signal segment; as well as The machine learning model is adjusted based on the target signal-only segment and the interference signal-only segment.

2. The method according to claim 1, characterized in that Classifying each input signal segment additionally includes: estimating a target signal-to-interference ratio of the input signal segment; categorizing the input signal segment as a target signal-only segment in response to the target signal-to-interference ratio exceeding a first threshold; and The input signal segment is classified as an interference-only signal segment in response to the target signal-to-interference ratio being less than a second threshold that is lower than the first threshold.

3. A method according to any of the preceding claims, characterised in that Adjusting the machine learning model further includes: A plurality of target-plus-interference signal segments are generated by mixing a selected target-only signal segment with a selected interference-only signal segment at a target signal-to-interference ratio to obtain each target-plus-interference signal segment.

4. The method according to claim 1, characterized in that: The input signal further comprises an audio signal and a video signal, and wherein classifying each input signal segment further comprises: determining noise power based on the audio component of the input signal segment; responsive to the noise power being less than a noise power threshold and voice activity being detected from a video component of the input signal segment, classifying the input signal segment as a target-only signal segment; and The input signal segment is classified as an interference-only signal segment in response to the noise power being greater than the noise power threshold and no voice activity being detected from the video component of the input signal segment.

5. The method according to claim 4, characterized in that Adjusting the machine learning model further includes: generating multiple target plus interference signal segments; wherein the audio component of the target plus interference signal segment is generated by mixing the audio component of the selected target-only signal segment with the audio component of the selected interference-only signal segment at a target signal-to-interference ratio; and The video component of the target plus interference signal segment includes the video component of the selected target-only signal segment.

6. The method according to claim 3 or 5, characterized in that: Adjusting the machine learning model further includes: constructing a target plus interference data set from the plurality of target plus interference signal segments; constructing a target signal-only data set from the corresponding selected target signal-only segments of each target-plus-interference signal segment; and The target plus interference dataset and the target signal only dataset are applied to the machine learning model.

7. The method according to claim 1, characterized in that The input signal comprises a digital audio signal, wherein each input signal segment comprises an audio packet, the audio packet comprising a plurality of audio samples, and wherein classifying each input signal segment further comprises: categorizing the input signal segments as only target signal segments in response to correctly receiving the audio packets; and The input signal segment is classified as an interference-only signal segment in response to incorrectly receiving the audio packet.

8. The method according to claim 7, characterized in that Also includes: At least one parameter of a packet loss pattern model is updated using the interference-only signal segment.

9. The method according to claim 8, characterized in that Adjusting the machine learning model further includes: constructing a target signal-only dataset from a plurality of selected target signal-only segments; constructing a target plus interference data set by muting a subset of the plurality of selected target-only signal segments according to a pattern loss packet model; and The target plus interference dataset and the target signal only dataset are applied to the machine learning model.

10. An interference signal reduction system, characterized in that: include: a signal input configured to receive an input signal comprising a plurality of signal segments; Interference signals reduce system output; a target signal estimator configured to output an estimate of a target signal based on the input signal, the target signal estimator comprising a machine learning model and having a target signal estimator input coupled to the signal input, a target signal estimator output coupled to the interference signal reduction system output, and an adapter input; a signal classifier having a signal classifier input coupled to the signal input, a target signal output, and an interference signal output; an adapter having a target signal input coupled to the target signal output, a jammer signal input coupled to the jammer signal output, and a model adapter output coupled to the adapter input; wherein the signal classifier is configured to classify each input signal segment of the plurality of signal segments into one of a target signal segment only, an interfering signal segment only, and optionally an undefined signal segment; and The adapter is configured to adjust the machine learning model based on the target-only signal segment and the interference-only signal segment.

Images