Introduction
1. This office action is in response to Applicant’s submission filed on 5/19/2026. Claims 22-41 are pending in the application and have been examined.
Notice of Pre-AIA or AIA Status
2. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
3. The Amendment submitted on 12/17/2025 has been entered and fully considered. With respect to the rejections of Claims 22-38 under 35 USC 103, the argument that the reference does not describe “output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers” is not persuasive. Figure 1 of Grais shows that the output of each layer is fed into the next layer, and thus eventually to every subsequent layer. The claim does not presently recite that the output of each layer is directly fed into the input of each subsequent layer, which is the scope of the arguments presented.
The rest of the arguments presented are moot in view of the following new rejections based on “WAVENET: A GENERATIVE MODEL FOR RAW AUDIO” (van den Oord et al., hereinafter “Oord”).
With respect to the rejection of Claims 39-41 under 35 USC 103, Applicant’s representative noted that the Zhang reference is not prior art with respect to the present application. Thus, the finality of the previous rejection is withdrawn and Claims 39-41 are newly rejected in part based on Oord, as noted below. Accordingly, the present rejection is a non-final rejection.
Claim Rejections - 35 USC § 103
4. In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
5. Claims 22, 23, 27-34, and 36-38 are rejected under 35 U.S.C. 103 as unpatentable over “Learning Complex Spectral Mapping With Gated Convolutional Recurrent Networks for Monaural Speech Enhancement” (Tan et al., hereinafter “Tan,” cited in IDS of 12/20/2023) in view of “SINGLE CHANNEL AUDIO SOURCE SEPARATION USING CONVOLUTIONAL DENOISING AUTOENCODERS” (Grais et al., hereinafter “Grais”), U.S. Pat. App. Pub. No. 20200219524 (Braun et al., hereinafter “Braun”), and “WAVENET: A GENERATIVE MODEL FOR RAW AUDIO” (van den Oord et al., hereinafter “Oord”).
With regard to Claim 22, Tan describes:
“A method of suppressing noise and enhancing speech, comprising:
receiving, by a processor, input audio data covering a plurality of frequency bands along a frequency dimension at a plurality of frames along a time dimension; (Section III(D), pg 384)
training, by the processor, a neural network model using the input audio data (Figure 5, Table 1, and Section IV(A), page 384), the neural network model comprising:
a feature extraction block that implements a lookahead of a specific number of frames in extracting features from the input audio data; (Table I and Section III(D) describe that timeSteps of frames are looked ahead.)
an encoder that includes a first series of blocks producing first feature maps corresponding to increasingly larger receptive fields in the input audio data along the frequency dimension; (Table I, page 384 describes that a number of channels decreases through each ConvGLU block (Figure 5), which leads to increasingly large receptor fields.)
a decoder that includes a second series of blocks receiving output feature maps generated by the encoder as input feature maps and producing second feature maps; (Figure 5 shows DeconvGLU blocks that receive the feature maps from the encoder and form second feature maps.)
a classification block that receives the second feature maps and generates a speech value indicating an amount of speech present for each frequency band of the plurality of frequency bands at each frame of the plurality of frames; (Figure 5 shows a “linear layer” that is described in Section IIID and Table I that maps the outputs to spectragrams.)
receiving new audio data comprising one or more frames; (Section IVB describes that a model is trained and then more data is input.)
executing the neural network model on the new audio data to generate new speech values for each frequency band of the plurality of frequency bands at each frame of the one or more frames; (Section IVB describes that the new audio data is used to generate new spectrograms using the trained model.)
transmitting the new output data. (Section V, pages 385-6 describes the output data.)
Tan does not explicitly describe:
“wherein each block of the first series of blocks comprising a feature computation block and a frequency down-sampler, the feature computation block comprising a series of convolutional layers, and wherein output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers, the series of convolutional layers implementing increasingly large dilation along the time dimension;” or
“generating new output data suppressing noise in the new audio data based on the new speech values by applying the new speech values generated by the neural network model to the new audio data.”
However, Grais describes:
“wherein each block of the first series of blocks comprising a feature computation block and a frequency down-sampler (Figure 1 shows ConvReLU blocks paired with pooling blocks that down sample), the feature computation block comprising a series of convolutional layers, and wherein output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers. (Figure 1 shows that the output of each convolutional layer is fed into subsequent layers, and thus indirectly into all subsequent layers)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the first series of blocks as described by Grais into the invention of Tan to suppress noise when determining sound from multiple sources, as described in section 3 of Grais.
Tan in view of Grais does not explicitly describe “generating new output data suppressing noise in the new audio data based on the new speech values by applying the new speech values generated by the neural network model to the new audio data” or “the series of convolutional layers implementing increasingly large dilation along the time dimension.”
However, Braun describes “generating new output data suppressing noise in the new audio data based on the new speech values by applying the new speech values generated by the neural network model to the new audio data.”
Paragraph 21 of Braun describes that noise reduction can be performed on new audio data by modifying that data with the output of a model. In particular, “by estimating the coefficients of the autoregressive reverberation model associated with a certain frame on the basis of a delayed and noise reduced reverberant signal which may be associated with one or more preceding frames, and that it is possible to provide the noise reduced reverberant signal of the current frame using the input audio signal and the estimated coefficients of the autoregressive reverberation model associated with the current frame and obtained on the basis of noise-reduced (and typically reverberant) signals (for example, provided by the noise reduction stage) associated with one or more preceding frames.”
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the modification of new audio data with the output of a noise reduction model as described by Braun into the invention of Tan in view of Grais to reduce the complexity of the computations used, as described in paragraph 21 of Braun.
Tan in view of Grais and Braun does not explicitly describe “the series of convolutional layers implementing increasingly large dilation along the time dimension.”
However, Figure 3 of Oord and the related description presents a series of convolutional layers implementing increasingly large dilation along the time dimension.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the layer dilation as described by Oord into the invention of Tan in view of Grais and Braun to enable networks to have very large receptive fields with just a few layers, as described in section 2.1 of Oord.
With respect to Claim 23, Tan describes:
“receiving an input waveform; (Figures 2 and 5 show input waveform Y)
transforming the input waveform into raw audio data covering a plurality of frequency bins along the frequency dimension at the one or more frames along the time dimension; (Table 1 shows the transformation parameters for converting to frequency domain in the encoder.)
converting the raw audio data into the new audio data by grouping the plurality of frequency bins into the plurality of frequency bands; (Table 1 shows the transformation parameters for frequency domain calculations.)
performing inverse banding on the new speech values to generate updated speech values for each frequency bin of the plurality of frequency bins at each frame of the one or more frames; (Table 1 shows the transformation parameters used in the decoder to map the data to spectrograms.)
applying the updated speech values to the raw audio data to generate the new output data; (Section IVB describes that the new audio data is used to generate new spectrograms using the trained model.)
transforming the new output data into an enhanced waveform. (Section V describes the noise suppression performance of the model.)
With regard to Claim 25, Tan describes “the feature extraction block comprises a convolution kernel that has a specific size along the time dimension (Page 386, 1st full paragraph describes that 4 different kernel sizes are used.), the specific size being larger than a size along the time dimension of any convolution kernel in the encoder or the decoder.” (Table 1 shows that all of the kernel sizes on page 386 are larger than the kernel size in the encoder and decoder.)
With regard to Claim 26, Tan describes “the feature extraction block comprises a batch normalization layer followed by a convolutional layer with a two-dimensional convolution kernel. (Table 1 and the last paragraph of Section III describe the use of a batch normalization layer and a 2-dimensional kernel.)
With regard to Claim 27, Tan describes “each of the feature extraction block, the first series of blocks, and the second series of blocks produce a common number of feature maps.” (Table 1 shows the dimensions of the series of block and that they produce a common number of feature maps.)
With regard to Claim 28, Tan describes “each of the series of convolutional layers comprises depth-wise separable convolutional blocks with a gating mechanism.” (Second last paragraph of Section I, page 381 describes that each of the convolutional layers is a gated linear unit.)
With regard to Claim 29, Tan describes “each of the series of convolutional layers comprises a residual block having a series of convolutional blocks, including a first convolutional block having a first one-by-one two-dimensional convolution kernel and a last convolutional block having a last one-by-one two-dimensional convolution kernel.” (Table 1 shows that the layers include the conv_2d blocks (“a residual block”) with a first convolutional block having a first one-by-one two-dimensional convolution kernel and a last convolutional block having a last one-by-one two-dimensional convolution kernel.)
With regard to Claim 30, Tan describes “output data of a feature computation block in a block of the first series of blocks is scaled by a learnable weight to form scaled output data, and wherein the scaled output data is communicated to a block of the second series of blocks in the decoder via a skip connection.” (Section IIIB describes the calculations that occur before the skip connection. Weighting σ scales the data. Figure 5 shows the use of skip connections.)
With regard to Claim 31, Tan describes “wherein a [[frequency down-sampler of a]] block in the first series of blocks comprises convolution kernels with a stride size greater than one along the frequency dimension.” (Table 1 shows that the encoder has blocks with a stride of 2.)
Tan does not explicitly describe down-sampler blocks. However, Figure 5 of Grais shows down-sampling blocks.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the down-sampling blocks as described by Grais into the invention of Tan to help suppress noise when determining sound from multiple sources, as described in section 3 of Grais.
With regard to Claim 32, Tan does not explicitly describe this feature. However, Grais describes “each block of the second series of blocks comprises a feature computation block and a frequency up-sampler.” Figure 5 and the related description of Grais shows a series of blocks with a computation block and an up-sampler.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the up-sampling blocks as described by Grais into the invention of Tan to help suppress noise when determining sound from multiple sources, as described in section 3 of Grais.
With regard to Claim 33, Tan describes “a feature computation block in a block of the second series of blocks receiving first output data from a feature computation block in a block of the first series of blocks and second output data from [[a frequency up-sampler of]] a previous block in the second series of blocks (Figure 5 shows the computation blocks receiving output from previous blocks), the first output data and the second output data being concatenated or added to form specific input data for the feature computation block in the block of the second series of blocks. (Section IIIA describes that the outputs are concatenated by the skip connections in Figure 5.)
Tan does not explicitly describe the use of an up-sampler. However, Figure 5 and the related description of Grais shows a series of blocks with a computation block and an up-sampler.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the up-sampling blocks as described by Grais into the invention of Tan to help suppress noise when determining sound from multiple sources, as described in section 3 of Grais.
With regard to Claim 34, Tan describes “the classification block comprises a one-by-one two-dimensional convolution kernel and a nonlinear activation function.” Table 1 and the last paragraph of Section III describe the use of a one-by-one two-dimensional convolution kernel and a nonlinear activation function.
With regard to Claim 36, Tan describes “the classification block further generates a distribution of speech amounts over a frequency band of the plurality of frequency bands at a frame, with the speech value being a mean of the distribution.” Section VA describes that the results are an average of the output spectrograms.
With regard to Claim 37, Tan describes “the input audio data comprises data corresponding to speech of different speeds or emotions, data containing different levels of noise, or data corresponding to different frequency bins.” Section IVA describes that sound with different noise levels is used.
With regard to Claim 38, Tan describes “the neural network model further comprises a feature computation block being output data of the encoder and input data of the decoder.” Figure 5 shows a feature computation block between the encoder and the decoder.
6. Claim 24 is rejected under 35 U.S.C. 103 as unpatentable over Tan in view of Grais, Braun, and Oord further in view of U.S. Pat. App. Pub. No. 20140310011 (Biswas et al., hereinafter “Biswas”).
With respect to Claim 24, Tan in view of Grais, Braun, and Oord does not explicitly describe this subject matter. However, Biswas describes “the plurality of frequency bands comprise perceptually motivated bands, covering more frequency bins at higher frequencies.” Paragraph 89 describes the used of perceptually motivated bands.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the perceptually motivated bands as described by Biswas into the invention of Tan in view of Grais, Braun, and Oord to further suppress noise, as described in paragraph 89 of Biswas.
7. Claim 35 is rejected under 35 U.S.C. 103 as unpatentable over Tan in view of Grais, Braun, and Oord and further in view of U.S. Pat. App. Pub. No. 20200312345 (Fazeli et al., hereinafter “Fazeli”).
With regard to Claim 35, Tan in view of Grais, Braun, and Oord does not explicitly describe this subject matter. However, paragraph 74 of Fazeli describes the claimed training procedure of “the training is performed with a function of loss between a predicted speech value and a ground-truth speech value for each frequency band of the plurality of frequency bands at each frame, with a larger weight in the function of loss when the predicted speech value corresponds to over-suppression of speech and a smaller weight in the function of loss when the predicted speech value corresponds to under-suppression of speech.”
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the training procedure as described by Fazeli into the invention of Tan in view of Grais, Braun, and Oord to efficiently train the model, as described in paragraph 89 of Fazeli.
8. Claims 39-41 are rejected under 35 U.S.C. 103 as unpatentable over Tan in view of Grais and further in view of Oord.
With regard to Claim 39, Tan describes “A system, comprising:
a memory; (Section V, bottom of page 386 describes that the method is bother processor and memory efficient.)
one or more processors coupled with the memory and configured to perform: (Section V, bottom of page 386 describes that the method is bother processor and memory efficient.)
receiving input audio data covering a plurality of frequency bands along a frequency dimension at a plurality of frames along a time dimension; (Section III(D), pg 384)
training a neural network model using the input audio data (Figure 5, Table 1, and Section IV(A), page 384), the neural network model comprising:
a feature extraction block that implements a lookahead of a specific number of frames in extracting features from the input audio data; (Table I and Section III(D) describe that timeSteps of frames are looked ahead.)
an encoder that includes a first series of blocks producing first feature maps corresponding to increasingly larger receptive fields in the input audio data along the frequency dimension; (Table I, page 384 describes that a number of channels decreases through each ConvGLU block (Figure 5), which leads to increasingly large receptor fields.)
a decoder that includes a second series of blocks receiving output feature maps generated by the encoder as input feature maps and producing second feature maps; (Figure 5 shows DeconvGLU blocks that receive the feature maps from the encoder and form second feature maps.)
a classification block that receives the second feature maps and generates a speech value indicating an amount of speech present for each frequency band of the plurality of frequency bands at each frame of the plurality of frames; (Figure 5 shows a “linear layer” that is described in Section IIID and Table I that maps the outputs to spectragrams.)
storing the neural network model. (Section IV, page 385 describes storing the trained model.)
Tan does not explicitly describe:
“wherein each block of the first series of blocks comprises a feature computation block and a frequency down-sampler, the feature computation block comprising a series of convolutional layers, and wherein output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers, the series of convolutional layers implementing increasingly large dilation factors for each convolutional layer along the time dimension.”
However, Grais describes:
“wherein each block of the first series of blocks comprises a feature computation block and a frequency down-sampler (Figure 1 shows ConvReLU blocks paired with pooling blocks that down sample), the feature computation block comprising a series of convolutional layers, and wherein output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers (Figure 1 shows that the output of each convolutional layer is fed into subsequent layers, and thus indirectly into each of the subsequent layers.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the first series of blocks as described by Grais into the invention of Tan to suppress noise when determining sound from multiple sources, as described in section 3 of Grais.
Tan in view of Grais does not explicitly describe “the series of convolutional layers implementing increasingly large dilation factors for each convolutional layer along the time dimension.”
However, Figure 3 of Oord and the related description presents a series of convolutional layers implementing increasingly large dilation factors along the time dimension.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the layer dilation as described by Oord into the invention of Tan in view of Grais to enable networks to have very large receptive fields with just a few layers, as described in section 2.1 of Oord.
With regard to Claim 40, Tan describes “A method of suppressing noise and enhancing speech, comprising:
receiving, by a processor, new audio data comprising one or more frames; (Section III(D), pg 384)
executing, by the processor, a trained neural network model on the new audio data, wherein the neural network model has been trained to generate new speech values for each frequency band of a plurality of frequency bands at each frame of the one or more frames given the new audio data, (Section V describes the noise suppression performance of the model.)
the trained neural network model comprising computer-executable instructions for:
feature extraction block that implements a lookahead of a specific number of frames in extracting features from the new input audio data; (Table I and Section III(D) describe that timeSteps of frames are looked ahead.)
an encoder that includes a first series of blocks producing first feature maps corresponding to increasingly larger receptive fields in the input new audio data along the frequency dimension; (Table I, page 384 describes that a number of channels decreases through each ConvGLU block (Figure 5), which leads to increasingly large receptor fields.)
a computation block that connects the encoder and a decoder; (Figure 5 shows a block between the encoder and decoder.)
the decoder that includes a second series of blocks receiving output feature maps generated by the encoder as input feature maps and producing second feature maps; (Figure 5 shows DeconvGLU blocks that receive the feature maps from the encoder and form second feature maps.)
a classification block that receives the second feature maps and generates a speech value indicating an amount of speech present for each frequency band of the plurality of frequency bands at each frame of a plurality of frames; (Figure 5 shows a “linear layer” that is described in Section IIID and Table I that maps the outputs to spectragrams.) and
the neural network model being trained with the input audio data covering the plurality of frequency bands along a frequency dimension at the plurality of frames along a time dimension; (Section IVB describes that the model is trained using initial data.)
generating new output data suppressing noise in the new audio data based on the new speech values; (Section V describes the noise suppression performance of the trained model.)
transmitting the new output data. (Section V, pages 385-6 describes the output data.)
Tan does not explicitly describe:
“wherein each block of the first series of blocks comprising a feature computation block and a frequency down-sampler, the feature computation block comprising a series of convolutional layers, and wherein output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers, the series of convolutional layers implementing increasingly large dilation factors for each convolutional layer along the time dimension.”
However, Grais describes:
“wherein each block of the first series of blocks comprising a feature computation block and a frequency down-sampler (Figure 1 shows ConvReLU blocks paired with pooling blocks that down sample), the feature computation block comprising a series of convolutional layers, and wherein output data of a convolutional layer of the series of convolutional layers being fed into all subsequent convolutional layers of the series of convolutional layers (Figure 1 shows that the output of each convolutional layer is fed into subsequent layers, and thus indirectly into each subsequent layer.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the first series of blocks as described by Grais into the invention of Tan to suppress noise when determining sound from multiple sources, as described in section 3 of Grais.
Tan in view of Grais does not explicitly describe “the series of convolutional layers implementing increasingly large dilation factors for each convolutional layer along the time dimension.”
However, Figure 3 of Oord and the related description presents a series of convolutional layers implementing increasingly large dilation factors along the time dimension.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the layer dilation as described by Oord into the invention of Tan in view of Grais to enable networks to have very large receptive fields with just a few layers, as described in section 2.1 of Oord.
With regard to Claim 41, Tan in view of Grais does not explicitly describe this subject matter. However, Oord describes “the increasingly large dilation factors for each convolutional layer follow an exponential increase from one convolutional layer to a subsequent convolutional layer.”
However, Figure 3 of Oord and the related description presents a series of convolutional layers implementing exponentially increasingly large dilation factors along the time dimension.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the exponential layer dilation as described by Oord into the invention of Tan in view of Grais to enable networks to have very large receptive fields with just a few layers, as described in section 2.1 of Oord.
Conclusion
9. The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
U.S. Pat. No. 11,508,388 (Souden et al.) also describes using increasing dilation factors between convolutional layers.
10. Any inquiry concerning this communication or earlier communications from the examiner should be directed to EDWARD TRACY whose telephone number is (571)272-8332. The examiner can normally be reached Monday-Friday 9 AM- 5PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Bhavesh Mehta can be reached on 571-272-7453. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/EDWARD TRACY JR./Examiner, Art Unit 2656
/BHAVESH M MEHTA/Supervisory Patent Examiner, Art Unit 2656
Prosecution Insights