Notice of Pre-AIA or AIA Status
1. 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 Arguments
2. Applicant’s arguments with respect to claims 1 and 11 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Specifically, newly added limitations to claims 1, and claim 11 and 20 are taught by newly cited Vygon et al. “Learning Efficient Representations for Keyword Spotting with Triplet Loss,” herein Vygon and McDuff (US 2020/0279553).
Applicant’s arguments and amendments with respect to rejections made under 35 USC § 101 have been considered and are persuasive. The rejection made under 35 USC § 101 is withdrawn.
Claim Rejections - 35 USC § 103
3. 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 (i.e., changing from AIA to pre-AIA ) 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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
4. Claims 1, 9, 10-11 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over San Martin (US 2015/0127594) in view of Vygon et al. “Learning Efficient Representations for Keyword Spotting with Triplet Loss,” herein Vygon and further in view of McDuff (2020/0279553).
Regarding Claim 1:
San Martin discloses the method for training an artificial intelligence (Al) model, the method comprising:
generating a first keyword dataset and a second keyword dataset, pre-training the Al model based on the first keyword dataset (San Martin: Fig. 7, ¶[0017]-[0018] discloses an apparatus with a deep neural network that is trained on a first and second training set of feature vectors); and
refining the pre-trained Al model through training with the second keyword dataset (San Martin: ¶[0017]-[0018] discloses after training on words/sub word units , it then trains with a second, smaller training set for specific keywords and key phrases),
wherein a number of keywords included in the first keyword dataset is greater than a number of keywords included in the second keyword dataset(San Martin: ¶[0004] ¶[0035]-[0039] discloses the first training set is much broader and larger than the second set, the first is used for general speech recognition, the second training set for specific keywords plus filler/negative examples. The first training set is a keyword dataset because it includes feature vectors modeling words or sub-word units of speech, while the second training set is a narrower keyword dataset directed to specific target keywords/key phrases. San Martin expressly teaches the second training set is smaller and more specific than the first, meaning the number of keywords would be less),
San Martin does not explicitly disclose:
wherein the pre-training comprises performing metric learning on the Al model,
and wherein generating the first keyword dataset comprises removing one or more most frequently appearing words from filtered words.
However, Vygon discloses:
wherein the pre-training comprises performing metric learning on the Al model (Vygon: Sections 1.2 and 1.3 discloses keyword spotting models trained using triplet loss based metric embeddings, explicitly defining metric embedding learning as learning a representation in which semantically similar samples are metrically close and semantically different samples are metrically distant).
San Martin and Vygon are combinable because they are from the same field of endeavor, i.e., both disclose methods for training neural network models for speech recognition and keyword spotting. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin to perform metric learning during pre-training using Vygon’s triplet loss based metric embedding technique. Such a modification would have predictably improved keyword classification accuracy because Vygon teaches: “triplet loss-based embedding and a variant of kNN for classification instead of cross-entropy loss significantly (by 26% to 38%) improves the classification accuracy for convolutional networks on a LibriSpeech- derived LibriWords datasets” within the abstract.
The proposed combination of San Martin in view of Vygon does not explicitly disclose
and wherein generating the first keyword dataset comprises removing one or more most frequently appearing words from filtered words.
However, McDuff discloses:
and wherein generating the first keyword dataset comprises removing one or more most frequently appearing words from filtered words (McDuff: ¶33, teaches that stop words are filtered out before or after processing natural language input, including words such as, a, the, is and in).
San Martin and Vygon in view of McDuff are combinable because they are from the same field of endeavor, i.e., each disclose systems or methods for processing speech and natural language processing word data using machine learning techniques. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin and Vygon to remove one or more of the most frequently appearing words from the filtered candidate words when generating the first keyword dataset as taught by McDuff. This modification would have predictably improved the resulting word set because McDuff teaches that “Stop words usually refers to the most common words in a language.”
Regarding Claim 9:
The proposed combination of San Martin, Vygon and McDuff the method of claim 1, wherein pre-training the AI model based on the first keyword dataset comprises configuring a batch using data from the first keyword dataset (San Martin: ¶[0003] and ¶[0017] discloses pre training the deep neural network using a first training set of feature vectors and further teaches estimating model parameters over labeled training data using asynchronous stochastic gradient descent in ¶[0034]. This is necessarily configured in a batch of training data from the first training set of iterative parameter updates, additionally,).
Regarding Claim 10
The proposed combination of San Martin, Vygon and McDuff the method of claim 9, wherein the batch includes at least one positive pair and at least one negative pair, the at least one positive pair including distinct audio data sharing a same keyword, and the at least one negative pair including audio data with different keywords (San Martin: ¶[0038]-[0039], ¶[0074] and ¶[0098] discloses keyword examples, negative examples. The second training set may include examples of the keywords or key phrases spoken by a particular user as positive examples and keywords or key phrases spoken by different users as negative examples, Vygon: Section 3.2 explicitly disclose triplet batch training examples).
San Martin and Vygon are combinable because they are from the same field of endeavor, i.e., both disclose methods for training neural network models for speech recognition and keyword spotting. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin to perform metric learning during pre-training using Vygon’s triplet loss based metric embedding technique. Such a modification would have predictably improved keyword classification accuracy because Vygon teaches: “triplet loss-based embedding and a variant of kNN for classification instead of cross-entropy loss significantly (by 26% to 38%) improves the classification accuracy for convolutional networks on a LibriSpeech- derived LibriWords datasets” within the abstract.
Regarding Claim 11:
San Martin discloses an apparatus configured to identify speech based on an artificial intelligence (AI) model, the apparatus comprising:
a dataset generator configured to generate a first keyword dataset and a second keyword dataset; and a model trainer, implemented using one or more computing devices (San Martin: ¶[0017]-[0018] discloses a deep neural network that is trained on a first training set of generated feature vectors),
configured to (i) pre-train an AI model based on the first keyword dataset and (ii) refine the pre-trained AI model through training with the second keyword dataset (San Martin: ¶[0017]-[0018] discloses after training on words/sub word units , it then trains with a second, potentially smaller training set for specific keywords and key phrases),
wherein a number of keywords included in the first keyword dataset is greater than a number of keywords included in the second keyword dataset (San Martin: ¶[0004] ¶[0035]-[0039] discloses the first training set is much broader and larger than the second set, the first is used for general speech recognition, the second training set for specific keywords plus filler/negative examples. The first training set is a keyword dataset because it include feature vectors modeling words or sub-word units of speech, while the second training set is a narrower keyword dataset directed to specific target keywords/key phrases. San Martin expressly teaches the second training set is smaller and more specific than the first, meaning the number of keywords would be less).
Regarding Claim 19:
Claim 19 has been analyzed with respect to claim 9 and 10 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Regarding Claim 20:
San Martin discloses an apparatus configured to identify a word from speech, the apparatus comprising:
a feature extractor configured to generate a feature vector from the speech (San Martin: Fig. 7, ¶[0017]-[0018] discloses an apparatus with a deep neural network that is trained on a first and second training set of feature vectors);
and an artificial intelligence (AI) model configured to output the word from the feature vector, wherein the AI model is pre-trained based on a first keyword dataset and is further refined through training with a second keyword dataset (San Martin: ¶[0017]-[0018] discloses after training on words/sub word units, it then trains with a second, potentially smaller training set for specific keywords and key phrases),
and a number of keywords included in the first keyword dataset is greater than a number of keywords included in the second keyword dataset (San Martin: ¶[0004] ¶[0035]-[0039] discloses the first training set is much broader and larger than the second set, the first is used for general speech recognition, the second training set for specific keywords plus filler/negative examples. The first training set is a keyword dataset because it include feature vectors modeling words or sub-word units of speech, while the second training set is a narrower keyword dataset directed to specific target keywords/key phrases. San Martin expressly teaches the second training set is smaller and more specific than the first, meaning the number of keywords would be less).
San Martin does not explicitly disclose:
wherein the AI model is pre-trained by performing metric learning using the first keyword dataset, and
wherein the first keyword dataset is generated by removing one or more most frequently appearing words from filtered words.
However, Vygon discloses:
wherein the AI model is pre-trained by performing metric learning using the first keyword dataset (Vygon: Sections 1.2 and 1.3 discloses keyword spotting models trained using triplet loss based metric embeddings, explicitly defining metric embedding learning as learning a representation in which semantically similar samples are metrically close and semantically different samples are metrically distant).
San Martin and Vygon are combinable because they are from the same field of endeavor, i.e., both disclose methods for training neural network models for speech recognition and keyword spotting. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin to perform metric learning during pre-training using Vygon’s triplet loss based metric embedding technique. Such a modification would have predictably improved keyword classification accuracy because Vygon teaches: “triplet loss-based embedding and a variant of kNN for classification instead of cross-entropy loss significantly (by 26% to 38%) improves the classification accuracy for convolutional networks on a LibriSpeech- derived LibriWords datasets” within the abstract.
The proposed combination of San Martin in view of Vygon does not explicitly disclose:
wherein the first keyword dataset is generated by removing one or more most frequently appearing words from filtered words.
However, McDuff discloses:
wherein the first keyword dataset is generated by removing one or more most frequently appearing words from filtered words (McDuff: ¶33, teaches that stop words are filtered out before or after processing natural language input, including words such as, a, the, is and in).
San Martin and Vygon in view of McDuff are combinable because they are from the same field of endeavor, i.e., each discloses systems or methods for processing speech and natural language processing speech and natural language word data using machine learning techniques. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin and Vygon to remove one or more of the most frequently appearing words from the filtered candidate words when generating the first keyword dataset as taught by McDuff. This modification would have predictably improved the resulting word set because McDuff teaches that “Stop words usually refers to the most common words in a language.”
5. Claims 2-4, 8, 12-14 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over San Martin (US 2015/0127594) in view Vygon, further in view of McDuff, further in view of Rao US (2021/0005184) and further in view of Kaushik (US 2022/0130384).
Regarding Claim 2:
San Martin further discloses the method of claim 1, wherein generating the first keyword dataset and the second keyword dataset (San Martin: ¶[0017]-[0018] discloses a deep neural network that is trained on a first training set of feature vectors) (Rao: ¶[0015], ¶[0038], ¶[0067]-[0070] and ¶[0089] discloses training data including utterance audio and transcription (plurality of words) being aligned by using a first CTC acoustic model to determine transcriptions).
San Martin and Rao are combinable because they are from the same field of endeavor, speech recognition systems using acoustic and text based models. Both disclose methods of using voice commands / hot-words for training a model to recognize speech. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to generate San Martin’s first and second keyword datasets using Rao’s alignment technique so that the words within the speech corpus are time aligned properly with the corresponding utterance audio for training. The suggestion/motivation for doing so is disclosed in Rao ¶[0003] “using many training examples, statistics generated from the approximate alignments can be used to determine a CD phone inventory that allows high accuracy of the CD acoustic model.”
Regarding Claim 3:
The proposed combination of San Martin and Rao further discloses the method of claim 2, wherein the artificial neural network-based feature extractor is configured to: output the aligned plurality of words including at least one of a preceding phoneme or a proceeding phoneme of the words (Rao: ¶[0015] discloses the first CTC model determines alignments, then a set of context dependent states is determined based on those alignments, ¶[0018] further explains that the context dependent states are triphones. A triphone necessarily includes a left phoneme, a central phoneme, and a right phoneme, hence the term triphone. Because triphones represent a phoneme in the context of surrounding phonemes, Rao reaches outputting aligned speech units including at least one preceding phoneme or proceeding phoneme).
San Martin and Rao are combinable because they are from the same field of endeavor, speech recognition systems using acoustic and text based models. Both disclose methods of using voice commands / hot-words for training a model to recognize speech. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to generate San Martin’s first and second keywords datasets using Rao’s alignment technique and phoneme breakdown. The suggestion/motivation for doing so is disclosed in Rao ¶[0003] “using many training examples, statistics generated from the approximate alignments can be used to determine a CD phone inventory that allows high accuracy of the CD acoustic model.” In other words, structured phonemes allows for accurate alignment and breakdown of the training data.
Regarding Claim 4:
The proposed combination of San Martin and Rao further discloses the method of claim 2, wherein the artificial neural network-based feature extractor is configured to output the aligned plurality of words including a margin of a predetermined length relative to the plurality of words within the first speech corpus (San Martin: ¶[0029] discloses stacking feature frames to add sufficient left and right context, ¶[0042] discloses a fixed time window, ¶[0060] discloses the stack of the frames depends on the length of the keyword, key phrase or acoustic unit. In total San Martin teaches that the feature extraction model outputs feature vectors using left and right context of adjacent feature vectors to create a larger feature vector, it stacks contiguous frames to add sufficient left and right context, such as ten recently received frames and thirty previously received frames, and that the stack of frames may include a selected number of frames. Rao: discloses aligning utterance audio with transcript derived phonetic sequence).
San Martin and Rao are combinable because they are from the same field of endeavor, speech recognition systems using acoustic and text based models. Both disclose methods of using voice commands / hot-words for training a model to recognize speech. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to generate San Martin’s first and second keyword datasets using Rao’s alignment technique so that the words within the speech corpus are time aligned properly with the corresponding utterance audio for training. The suggestion/motivation for doing so is disclosed in Rao ¶[0003] “using many training examples, statistics generated from the approximate alignments can be used to determine a CD phone inventory that allows high accuracy of the CD acoustic model.”
Regarding Claim 8:
The proposed combination of San Martin and Rao further discloses the method of claim 2, wherein generating the first keyword dataset and the second keyword dataset comprises: generating the second keyword dataset based on a plurality of classes within a second speech corpus, the second speech corpus including fewer classes compared to the first speech corpus (San Martin: ¶[0005], ¶[0036]discloses the first training set comprises uttered speech with output values relevant to large vocabulary dictation and may involve a system having approximately 14,000 states as compared to a smaller embedded system having approximately 2,000 states ¶[0018] and ¶[0035] further disclose that the second training set is directed to specific keywords and key phrases and that when training with the second training set the system may replace the original output layer with a new output layer having one node for each keyword or key phrase and optionally one node for filler, therefore teaching a second speech corpus training set having fewer classes than the first speech corpus training set).
Regarding Claim 12:
Claim 12 has been analyzed with respect to claim 3 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Regarding Claim 13
Claim 13 has been analyzed with respect to claim 3 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Regarding Claim 14:
Claim 14 has been analyzed with respect to claim 4 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Regarding Claim 18:
Claim 18 has been analyzed with respect to claim 8 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
6. Claims 5-7, 15-17 are rejected under 35 U.S.C. 103 as being unpatentable over San Martin in view of Vygon, further in view of McDuff, further in view of Rao and further in view of Kaushik (US 2022/0130384).
Regarding Claim 5:
The proposed combination of San Martin and Rao further discloses the method of claim 2 wherein generating the first keyword dataset and the second keyword dataset comprises: identifying a word from the aligned utterance data through the artificial neural network-based feature extractor, (Rao: ¶[0015] discloses determining approximate alignments between audio data and phonetic sequences corresponding to transcriptions. Further ¶[0038], ¶[0046] and ¶[0052] disclose aligning training data to obtain phonetic alignments and frames with phone labels. ¶[0067]-[0070] discloses the alignment data indicates alignment between the audio data and the transcriptions. ¶[0073] discloses determining candidate word sequences from the model outputs. Collectively, Rao teaches identifying a word from aligned utterance data).
San Martin and Rao are combinable because they are from the same field of endeavor, speech recognition systems using acoustic and text based models. Both disclose methods of using voice commands / hot-words for training a model to recognize speech. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to generate San Martin’s first and second keywords datasets using Rao’s alignment technique for determining a specific keyword within the aligned sequence. The suggestion/motivation for doing so is disclosed in Rao ¶[0003] “using many training examples, statistics generated from the approximate alignments can be used to determine a CD phone inventory that allows high accuracy of the CD acoustic model.”
The combination of San Martin and Rao does not explicitly disclose and determining the first keyword dataset by filtering the identified word, however Kaushik discloses:
determining the first keyword dataset by filtering the identified word (Kaushik: ¶[0008]-[0011] discloses that pronunciation augmentation is followed by pruning to reduce the possible number of phoneme sequence variations and to render output phonemes, ¶[0100]-[0103] discloses pruning by eliminating pronunciations that are too short or insufficiently similar, ¶[0104]-[0109] then discloses sample-based pruning using positive and negative samples. Therefore, determining a keyword dataset by filtering the identified word).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin and Rao in view of Kaushik to include filter/pruning of identified candidate word or phoneme sequences when generating the keyword dataset. These references are in the same field of endeavor, speech recognition and keyword spotting. The suggestion/motivation for doing so is disclosed in Kaushik ¶[0100] where it is explicitly taught that pruning preserves good pronunciations.
Regarding Claim 6:
The proposed combination of San Martin, Rao and Kaushik further discloses the method of claim 5, wherein determining the first keyword dataset by filtering the identified word comprises: performing the filtering based on an edit distance between the identified word and a target word (Kaushik: ¶[0014] and ¶[0104] discloses that filtering and pruning may be performed based on similarity between sequences where the similarity is measured using a fused distance metric based on Jaro-Winkler normalized distance and Damerau Levenshtein normalized distance. Because Damerau Levenshtein normalized distance is an edit distance metric Kaushik teaches performing the filtering based on an edit distance between an identified word and a target word).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin and Rao in view of Kaushik to include filter/pruning of identified candidate word or phoneme sequences when generating the keyword dataset. These references are in the same field of endeavor, speech recognition and keyword spotting. The suggestion/motivation for doing so is disclosed in Kaushik ¶[0100] where it is explicitly taught that pruning preserves good pronunciations.
Regarding Claim 7:
The proposed combination of San Martin, Rao and Kaushik further disclose the method of claim 6, wherein performing the filtering based on the edit distance comprises: filtering the identified word based on an allowable range of the edit distance determined by a number of characters of the identified word (Kaushik: ¶[0014] and ¶[0103] discloses eliminating phoneme sequences that are shorter than a threshold sequence length. Also, filtering is based on distance threshold using a fused distance metric).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify San Martin and Rao in view of Kaushik to include filter/pruning of identified candidate word or phoneme sequences when generating the keyword dataset. These references are in the same field of endeavor, speech recognition and keyword spotting. The suggestion/motivation for doing so is disclosed in Kaushik ¶[0100] where it is explicitly taught that pruning preserves good pronunciations.
Regarding Claim 15:
Claim 15 has been analyzed with respect to claim 5 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Regarding Claim 16:
Claim 16 has been analyzed with respect to claim 6 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Regarding Claim 17:
Claim 17 has been analyzed with respect to claim 7 (see rejection above) and is rejected for the same reasons of obviousness set forth above.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to IAN SCOTT MCLEAN whose telephone number is (703)756-4599. The examiner can normally be reached "Monday - Friday 8:00-5:00 EST, off Every 2nd Friday".
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, Hai Phan can be reached at (571) 272-6338. 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.
/IAN SCOTT MCLEAN/Examiner, Art Unit 2654
/HAI PHAN/Supervisory Patent Examiner, Art Unit 2654
Prosecution Insights