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 filed have been fully considered but they are not persuasive. First, Applicant’s argument effectively requires the claimed spoken length value to be generated after and as a consequence of the speech-rate classification. This is not persuasive because it improperly imports this relationship into the claim. Rather, the claim requires a spoken length value “corresponding to the spoken keyword and the speech rate classification,” and information “indicative of a respective spoken length value of spoken keyword for each speech rate classification.” Tomashenko explicitly establishes the required relationship between speech duration and each speech-rate classification. Tomashenko calculates Rate of Speech (ROS) on a per-word basis using the actual duration of the triphones of the word relative to their expected mean durations, and then partitions the resulting duration based measure into slow, medium and fast classifications. Therefore, each speech rate classification is, by definition, associated with a respective range of normalized spoken durations. The fact that Tomashenko mathematically derives the classification from duration information rather than first classifying speech and thereafter selecting a duration does not distinguish the claim, which recites correspondence rather than a direction of calculation. Tomashenko, further explicitly associated each classification with a different temporal adaption. Fast: 7.5 ms, medium: 9.4 ms and slow: 10 ms. This confirms that each classification is associated with respective temporal information. Additionally, Applicant’s phrase “the opposite relationship” is misleading. Tomashenko’s relationship is mathematical, not one-way. For a given word, it uses triphone durations to calculate a normalized ROS value and defines duration domains. Once those thresholds are defined, knowing that a word falls into “slow” provides information about its normalized duration, just as the duration tells you which class it falls into. Applicant’s argument treats a classification function as if it destroys the information from which the classification was constructed.
Further, Applicant’s argument appears to require a separately stored final keyword length value for each speech rate classification. Claim 1 does not recite this. The claim requires information that is “indicative of” a respective spoken length value for each classification. Tomashenko defines each of the slow, medium and fast classifications by respective ranges of a duration based ROS measure derived from actual triphone durations and expected mean triphone durations. In combination with Jiang’s keyword pronunciation and constituent phoneme duration information, this information is indicative of a respective spoken duration of the keyword for each speech rate classification. Therefore, Applicant’s characterization of Tomashenko’s expected durations as “class-independent” does not eliminate the explicitly disclosed relationship between duration and each of the resulting speech-rate classifications.
Further, Applicant argues that Jiang discloses a minimal phoneme duration table not an average keyword length table. Applicant is correct in this assertion; however, Jiang still provides a known keyword and phoneme sequences, stored phoneme duration information, candidate phoneme segments with recorded start and end points and a second stage explicitly classed a refinement that uses phonemic duration “to define a word boundary and to refine and verify the search.” Jiang also loads the DNN , phoneme-duration table, triphone lookup table and keyword pronunciation table together during initialization. Tomashenko provides a rate-dependent duration relationship.
Moreover, Applicant’s argument improperly treats the 103 rejections as requiring Jiang to recite each mathematical operation verbatim. Jiang determines candidate start and end points, from which the candidate duration is determined and uses phonemic duration to define and refine the word boundary. In view of the Tomashenko’s rate dependent duration teachings, comparing that detected duration with the duration corresponding to the classified speech rate to determine the refined temporal boundaries is a predictable use of the combined teachings because of Jiang’s stated boundary refinement purpose.
The argument that Jiang merely provides pass or fail is not persuasive. Applicant’s characterization focuses on the later pass/fail pseudocode but does not account for Jiang’s explicit description of the preceding boundary refinement stage. Jiang explicitly teaches before that, the second stage uses phonemic duration “to define a word boundary and to refine and verify the search.” That is not merely verification it explicitly says refine and define a word boundary. Jiang records segment starting and end points and then applies average frames/phone and individual phoneme duration constraints before saving the keyword timing information.
In view of the arguments set forth above, the rejection of claims 1-9, 11-22, 24-27, 29 and 31-32 is maintained.
Examiner Comment
3. Prior Office actions cited Jiang using Column and line numbers from U.S. Patent No. 10,074,363. For clarity and ease of reference, the present Office action cites the corresponding pre-grant publication, U.S. Patent Application Publication No. US 2017/0133038, using paragraph numbers. The corresponding disclosure relied upon is the same. This change in citation format does not represent a change in the Jiang teachings relied upon in the 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 (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.
5. Claims 1-2, 4-9, 11-20, 22, 24-27, 29 and 31-32 are rejected under 35 U.S.C. 103 as being unpatentable over Jiang et al. (US 2017/0133038) (herein Jiang) in view of Tomashenko et al. “Speaking Rate Estimation Based on Deep Neural Networks” (herein Tomashenko).
Regarding Claim 1:
Jiang discloses an apparatus for processing one or more audio samples, comprising:
one or more memories configured to store the one or more audio samples (Jiang: ¶¶[0006]-[0007] discloses an apparatus with memory to store audio instructions for processing it);
and one or more processors coupled to the one or more memories (Jiang: ¶¶[0006]-[0007] discloses a processor to execute audio speech recognition), the one or more processors being configured to:
detect, using a first keyword detection model, a spoken keyword within an audio sample of the one or more audio samples (Jiang: ¶[0015] Jiang explicitly discloses it performs keyword spotting by matching keyword phoneme patterns to audio derived phoneme image maps);
determine estimated keyword indices corresponding to detection of the spoken keyword within the audio sample, the estimated keyword indices comprising an estimated keyword start index and an estimated keyword end index (Jiang: ¶[0015] discloses initial keyword candidates include estimated temporal start/end points, ¶[0027] The Online Keyword Speech Recognition workflow discloses explicitly recording the estimated start and end indices);
obtain, from information embedded in machine learning model metadata of the speech rate classification machine learning network, a spoken length value corresponding to the spoken keyword (Jiang: ¶[0027] The Online Keyword Speech Recognition workflow discloses a stored phoneme duration table, triphone lookup table and keyword pronunciation table loaded together with the DNN model, the known keyword pronunciation identified its constituent phonemes and the duration information provides the corresponding spoken length information. Jiang explicitly initializes the system by: Loading the DNN model Loading the triphone map; Loading the phoneme duration table; Load the triphone lookup table; Loading the keyword pronunciation table)
determine an estimated length for the spoken keyword, based on a difference between the estimated keyword end index and the estimated keyword start index (Jiang: ¶[0027] The Online Keyword Speech Recognition workflow and ¶[0028] Confidence Score Computation disclose logging the starting and ending points of the detected keyword and explicitly define the matched keyword as starting at t1 and ending at t2, therefore determining the temporal length extending from the start point to the end point. Jiang explicitly states: “finding consecutive frames and log the starting and end points” and “assume that the matched keyword starts at t1 and ends at t2”)
compare the estimated length to the spoken length value to determine a refined length for the spoken keyword (Jiang: ¶[0015] and ¶[0027] disclose comparing the detected keyword duration against phonemic duration information during the second, refinement stage to determine the proper word boundary, Jiang explicitly applies “Average frames/phone >= 6” and “individual phoneme duration constraints” in determining the refined keyword timing);
and generate refined keyword indices based on the refined length for the spoken keyword, the refined keyword indices include a refined keyword start index shifted to a time earlier than the estimated keyword start index (Jiang: ¶[0015] discloses using phonemic duration and acoustic information to define and refine the initially detected word boundary; Jiang further discloses merging neighboring phoneme segments having small gaps, whereby merging and earlier neighboring segment extends the detected segment to an earlier starting point and produces the refined keyword start index)
Jiang does not explicitly disclose:
determine, using a speech rate classification machine learning network, a speech rate classification corresponding to the audio sample, the speech rate classification included in a plurality of speech rate classifications associated with the speech rate classification machine learning network;
and the speech rate classification, wherein the information is indicative of a respective spoken length value of the spoken keyword for each speech rate classification of the plurality of speech rate classifications;
However, Tomashenko discloses
determine, using a speech rate classification machine learning network, a speech rate classification corresponding to the audio sample, the speech rate classification included in a plurality of speech rate classifications associated with the speech rate classification machine learning network (Tomashenko: Section 3 discloses a deep neural network trained to classify speech as slow, medium, fast or silence);
and the speech rate classification, wherein the information is indicative of a respective spoken length value of the spoken keyword for each speech rate classification of the plurality of speech rate classifications (Tomashenko: Section 2, discloses determining rate of speech as actual triphone duration relative to expected mean triphone duration, including on a per-word basis and mapping the resulting duration based ROS measure into respective slow, medium and fast classifications);
It would have been obvious for one of ordinary skill in the art before the affective filing date of the claimed invention to modify Jiang’s keyword boundary refinement process to incorporate the speaking-rate normalization taught in Tomashenko. Both are within the same field of endeavor, audio speech recognition and disclose methods for understanding keywords and phrases better within a user utterance. Jiang already relies on phoneme triphone average duration statistics to refine keyword start and end boundaries. Jiang already relies on phoneme/triphone duration information to refine keyword start and end boundaries. Tomashenko teaches that speaking rate is a significant source of speech variability and determines speaking rate using actual triphone durations relative to expected mean triphone durations, including on a per-word basis, with the resulting duration based measure classified as slow, medium or fast. The motivation for doing so is disclosed in Tomashenko’s Introduction “A reliable estimation of speaking rate is often needed in order to adapt the ASR system to variations in speaking rate.” Therefore, the combination yields no change in overall system operation and merely applies a known adjustment (rate normalization) to an existing duration-based boundary refinement method.
Regarding Claim 2:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, wherein the speech rate classification is indicative of a slow speech rate classification, a normal speech rate classification, or a fast speech rate classification for the spoken keyword within the audio sample (Tomashenko: Section 3 discloses training the DNN model to output slow, normal, fast. These outputs directly correspond to the speech rate information).
It would have been obvious for one of ordinary skill in the art before the affective filing date of the claimed invention to incorporate Tomashenko’s keyword speech rate classification into Jiang. Both are within the same field of endeavor, audio speech recognition and disclose methods for understanding keywords and phrases better within a user utterance. Tomashenko expressly classifies speech rate into slow, normal and fast categories. Claim 1 already recites determining speech rate which is taught using the model of Tomashenko, representing the speech rate information as one of these explicit categories is taught by the prior art already and would equate to a simple modification of Jiang’s base structure without undue experimentation. The motivation for doing so is disclosed in Tomashenko’s Introduction “A reliable estimation of speaking rate is often needed in order to adapt the ASR system to variations in speaking rate.”
Regarding Claim 4:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, wherein the one or more processors are configured to:
determine the speech rate classification using the speech rate classification machine learning network in response to detection of the spoken keyword (Tomashenko: Section 1 and 3 teaches computing speech rate for a target speech segment, including per word and per phone duration normalization using its overall machine learning architecture based on Deep learning networks and other formulas embedded into this process);
and determine the estimated keyword start index using a keyword start estimation neural network in response to detection of the spoken keyword (Jiang: ¶[0015] and ¶¶[0019]-[0020] explicitly performs refinement after a keyword has been detected which includes start and end boundary using phoneme durations and voicing cues using a deep neural network based acoustic triphone model).
It would have been obvious for one of ordinary skill in the art before the affective filing date of the claimed invention to modify Jiang’s keyword boundary refinement process to incorporate the speaking-rate normalization taught in Tomashenko. Both are within the same field of endeavor, audio speech recognition and disclose methods for understanding keywords and phrases better within a user utterance. Jiang already relies on phoneme triphone average duration statistics to refine keyword start and end boundaries. Jiang already relies on phoneme/triphone duration information to refine keyword start and end boundaries. Tomashenko teaches that speaking rate is a significant source of speech variability and determines speaking rate using actual triphone durations relative to expected mean triphone durations, including on a per-word basis, with the resulting duration based measure classified as slow, medium or fast. The motivation for doing so is disclosed in Tomashenko’s Introduction “A reliable estimation of speaking rate is often needed in order to adapt the ASR system to variations in speaking rate.” Therefore, the combination yields no change in overall system operation and merely applies a known adjustment (rate normalization) to an existing duration-based boundary refinement method.
Regarding Claim 5:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 4, wherein: the first keyword detection model and the keyword start estimation neural network are included in a first keyword detection stage of a multi-stage keyword detection system (Jiang: ¶¶[0015]-[0027] discloses a multistage keyword detection stage and a refinement stage that determines start and end boundaries).
Regarding Claim 6:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, wherein:
the first keyword detection model is configured to perform always-on keyword detection for one or more audio samples (Jiang: ¶[0038] receives an audio stream, under Broadest Reasonable Interpretation (BRI) this encompasses an always on mode as it is not limited to activation (e.g., pressing a button) because a stream is an ongoing audio from a microphone or live feed; ¶¶[0039]-[0041] discloses receiving the audio stream and processing the stream into phoneme frames and searching the audio stream image for keyword matches);
and the speech rate classification machine learning network is configured to perform speech rate classification for a particular audio sample of the one or more audio samples based on detection of the spoken keyword within the particular audio sample by the first keyword detection model (Tomashenko: Section 2-3 discloses a classifier that is not run continuously but applied to a selected segment which aligns with running it after keyword detection identifies the region).
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Jiang’s continuous keyword detection process such that, upon detection of a spoken keywork, the speech- rate classification taught by Tomashenko is performed for the audio corresponding to the detected keyword. Jiang and Tomashenko are in the same field of endeavor, i.e., speech processing and recognition. This modification would account for speech-rate variation at the particular detected word rather than requiring a single rate determination for an entire utterance. Tomashenko explicitly motivates such processing, stating “[i]n conversational speech, fluctuations in speech rate may be very high within one utterance, and per word ROS estimation may be more appropriate” in Section 2.
Regarding Claim 7:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, wherein:
the spoken length value is included in average keyword length information corresponding to the spoken keyword (Jiang: ¶[0020] discloses a stored phoneme duration table and stored pronunciations for each keyword, such that the pronunciation identifies the constituent phonemes of the particular keyword and the duration table provides corresponding duration information; Tomashenko: Section 2 further discloses expected mean triphone durations “estimated from the training database,” thereby providing average duration information for the speech units making up the words);
and the average keyword length information includes the respective average spoken length value for each speech rate classification the plurality of speech rate classifications associated with the speech rate classification machine learning network (Tomashenko: Section 2, discloses calculating speech rate on a per word basis using actual triphone durations relative to expected mean triphone durations and assigning the resulting duration based measure to respective slow, medium and fast speech rate classification. In combination with Jiang’s keyword specific pronunciation and duration information, the average keyword length information therefore includes duration information indicative of the respective average spoken length of the keyword for each speech rate classification).
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Jiang’s stored keyword duration information to incorporate the speech rate dependent duration relationship taught by Tomashenko. Jiang and Tomashenko are in the same field of endeavor, i.e., speech recognition using pgoonetic duration information. The modification would permit Jiang’s keyword duration information to account for speaking rate variability when refining keyword boundaries. Tomashenko explicitly explains that “of different phones and triphones in the database may vary a lot. To get more accurate estimations, we use normalization on the average length of triphones” in Section 2.
Regarding Claim 8:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 7, wherein the average keyword length information comprises offline estimations of the respective spoken length values (Jiang: ¶¶[0016]-[0019] discloses offline preparation mode; Tomashenko, Section 2 discloses that the expected mean triphone duration is estimated from the training database, therefore providing offline estimations of the duration information).
Regarding Claim 9:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 7, wherein the average keyword length information is included in the information embedded in the machine learning model metadata, and wherein the machine learning model metadata is associated with a configuration of the speech rate classification machine learning network or a configuration of a keyword indices refinement machine learning network used to generate the refined keyword indices (Jiang: ¶[0027] discloses initialization of the machine learning processing system by loading the deep neural network model (DNN) with the phoneme duration table, triphone map, triphone lookup take and keyword pronunciation table; Tomashenko: Sections 2-3 disclose the training derived expected duration information and the DNN speech-rate classification model)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Tomashenko’s trained duration information and speech rate classification parameters into the model associated configuration information loaded for Kiang’s keyword duration processing. Jiang explicitly initializes its machine learning processing by loading the DNN model together with the phoneme duration table, triphone map, triphone lookup table and keyword pronunciation table. Incorporating the speech rate information in Tomashenko into this model would make the information available to the machine learning processing during keyword boundary refinement. Tomashenko motivates this duration information stating that: “estimations, we use normalization on the average length of triphones, computed from the training database” in Section 2.
Regarding Claim 11:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, wherein, to generate the refined keyword indices, the one or more processors are configured to:
determine the refined keyword start index as a time index shifted earlier than the estimated keyword start index by a first amount corresponding to a difference between the refined length and the estimated length for the spoken keyword (Jiang: ¶[0015] discloses using phonemic duration and acoustic information to define a word boundary and to refine the initial search; ¶[0027] discloses logging initial starting and end points, connecting neighboring phoneme segments having small gaps and saving the resulting keyword timing information);
and determine a refined keyword end index as a time index shifted later than the estimated keyword end index by a second amount corresponding to the difference between the refined length and the estimated length for the spoken keyword (Jiang: ¶[0028] defines the matched keyword by its start t1 and end t2, therefore refinement of the detected temporal extent produces corresponding changes to the start and end indices based on the change in keyword length).
Regarding Claim 12:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 11, wherein the first amount and the second amount are the same (Jiang: ¶41 discloses temporal extension around a matched keyword, stating “fifteen seconds prior to the matched word and fifteen seconds after the identified word”).
Regarding Claim 13:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 11, wherein: the first amount comprises a first percentage of the difference between the refined length and the estimated length for the spoken keyword (Jiang: ¶[0041] discloses dividing a temporal extension of a detected keyword between an amount preceding the keyword);
and the second amount comprises a second percentage of the difference between the refined length and the estimated length for the spoken keyword (Jiang: ¶[0041] discloses an amount following the keyword, including an equal fifteen seconds allocation).
Regarding Claim 14:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 13, wherein the first percentage is greater than the second percentage (Jiang: ¶[0015] discloses refining the word boundary based on phonemic duration and notes that a vowel is longer; ¶[0020] discloses phoneme specific duration information in which long vowels are at least 5 frames while consonants are 2 frames; ¶[0027] applies individual phoneme duration constraints when refining the keyword timing, ¶[0041] discloses “[t]he amount of audio desired is up to individual preference”).
Regarding Claim 15:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 13, wherein the first percentage is greater than 50%, and wherein a sum of the first percentage and the second percentage is equal to 100% (Jiang: ¶[0020] discloses phoneme specific duration constraints, including long vowels having a minimum duration of 5 frames while consonants have a minimum duration of 2 frames; ¶[0027] discloses applying the individual phoneme duration constraints in refining the detected keyword timing, accordingly where the start side phoneme requires the greater duration adjustment, the start side portion constitutes more than half of the total boundary adjustment).
Regarding Claim 16:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, further comprising a microphone configured to obtain the one or more audio samples (Jiang: ¶[0045] and Fig. 5 user input devices (i.e., microphones)).
Regarding Claim 17:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, further comprising: one or more microphones configured to capture the one or more audio samples for keyword detection (Jiang: Jiang: ¶[0045] and Fig. 5 discloses microphones as input devices and processing received audio stream to identify keywords within the streams).
Regarding Claim 18:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 17, wherein the one or more microphones and the first keyword detection model are associated with an always-on keyword detection process implemented by the apparatus (Jiang: ¶[0039] discloses receiving an audio stream; ¶[0041] discloses continuously searching the audio file or stream representation for keyword matches and ¶¶[0045]-[0046] discloses microphones as system input devices and execution of the keyword identification process on received audio streams);
Regarding Claim 19:
Claim 19 has been analyzed with respect to claim 1 and is rejected for the same reasons of obvious as used above.
It is noted that the processors in Jiang used to teach the apparatus of claim 1 perform the method as disclosed in claim 19.
Regarding Claim 20:
Claim 20 has been analyzed with respect to claim 2 and is rejected for the same reasons of obvious as used above.
Regarding Claim 22:
Claim 22 has been analyzed with respect to claim 4 and is rejected for the same reasons of obvious as used above.
Regarding Claim 24:
Claim 24 has been analyzed with respect to claim 6 and is rejected for the same reasons of obvious as used above.
Regarding Claim 26:
Claim 26 has been analyzed with respect to claim 8 and is rejected for the same reasons of obvious as used above.
Regarding Claim 27:
Claim 27 has been analyzed with respect to claim 9 and is rejected for the same reasons of obvious as used above.
Regarding Claim 29:
Claim 29 has been analyzed with respect to claim 11 and is rejected for the same reasons of obvious as used above.
Regarding Claim 31:
The proposed combination of Jiang in view of Tomashenko further discloses the processor-implemented method of claim 19, wherein the spoken length value is an average spoken length value corresponding to the spoken keyword (Jiang: 4:18-32, Col 5 First table, discloses obtaining a baseline spoken length value corresponding to the spoken keyword through stored phoneme duration constraints and keyword pronunciation information, further discloses loading the phoneme duration table and the keyword pronunciation table for use in keyword boundary refinement) and the speech rate classification (Tomashenko discloses that speaking rate is classified into a plurality of speech rate classifications, therefore teaching an average spoken length value corresponding to the spoken keyword and the speech rate classification).
It would have been obvious for one of ordinary skill in the art before the affective filing date of the claimed invention to modify Jiang’s keyword boundary refinement process to incorporate the speaking-rate classification in Tomashenko. Both are within the same field of endeavor, audio speech recognition and disclose methods for understanding keywords and phrases better within a user utterance. Jiang already relies on phoneme triphone average duration statistics to refine keyword start and end boundaries. Tomashenko teaches classification according to measured speech rate (see Section 2 of Tomashenko). The motivation for doing so is disclosed in Tomashenko’s Introduction “A reliable estimation of speaking rate is often needed in order to adapt the ASR system to variations in speaking rate.” Therefore, the combination yields no change in overall system operation and merely applies a known adjustment (rate normalization) to an existing duration-based boundary refinement method.
Regarding Claim 32:
Claim 32 has been analyzed with respect to claim 31 and is rejected for the same reasons of obvious as used above.
6. Claims 3 and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Jiang, in view of Tomashenko and further in view of Parada San Martin (US 9,715,660).
Regarding Claim 3:
The proposed combination of Jiang and Tomashenko further discloses the apparatus of claim 1, wherein the one or more processors are configured to determine the speech rate classification (Tomashenko: Introduction Page 419, explicitly teaches a deep neural network (DNN) which classifies speech rate from acoustic feature and determine the estimated keyword start index Jiang: ¶[0015] discloses general search stage produces starting and ending key word boundaries)
Neither Jiang nor Tomashenko explicitly state that this process is done in parallel. However, Parada San Martin discloses processing in parallel (Parada San Martin: Col 17:60 - Col 18:14 discloses that speech/audio analysis systems may perform multiple detection/classification operations in parallel rather than in a strict sequence).
It would have been obvious for one of ordinary skill in the art to disclose processing in parallel for keyword detection systems. The combination of Jiang and Tomashenko are within the same field of endeavor as Parada San Martin, audio speech recognition of keywords. i.e., both disclose processing speech and audio to recognize key or hot-words. The motivation for doing so is “In certain circumstances, multitasking and parallel processing may be advantageous” as discloses by Parada San Martin in Col 17. This shows that depending on the use case parallel processing can be used to reduce processing delay, improving real-time performance and therefore would have been obvious to implement both processes happening in Jiang and Tomashenko in parallel.
Regarding Claim 21:
Claim 21 has been analyzed with respect to claim 3 and is rejected for the same reasons of obvious as used above.
Conclusion
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
