Ex Parte Hannun et al

Patent Trial and Appeal Board

Mar 28, 2019

14735002 (P.T.A.B. Mar. 28, 2019)

UNITED STA TES p A TENT AND TRADEMARK OFFICE
APPLICATION NO. FILING DATE
14/735,002 06/09/2015
119276 7590 04/01/2019
BAIDU USA LLC
c/o NORTH WEBER & BAUGH LLP
3260 Hillview A venue
Palo Alto, CA 94304
FIRST NAMED INVENTOR
Awni Hannun
UNITED STATES DEPARTMENT OF COMMERCE
United States Patent and Trademark Office
Address: COMMISSIONER FOR PATENTS
P.O. Box 1450
Alexandria, Virginia 22313-1450
www .uspto.gov
ATTORNEY DOCKET NO. CONFIRMATION NO.
28888-1910 3755
EXAMINER
SARPONG, AKWASI
ART UNIT PAPER NUMBER
2675
NOTIFICATION DATE DELIVERY MODE
04/01/2019 ELECTRONIC
Please find below and/or attached an Office communication concerning this application or proceeding.
The time period for reply, if any, is set in the attached communication.
Notice of the Office communication was sent electronically on above-indicated "Notification Date" to the
following e-mail address(es):
docket!@northweber.com
docket2@northweber.com
bbaugh@northweber.com
PTOL-90A (Rev. 04/07)
UNITED STATES PATENT AND TRADEMARK OFFICE
BEFORE THE PATENT TRIAL AND APPEAL BOARD
Ex parte LEE LINDEN, BENJAMIN LEWIS, and ABHEEK ANAND
Appeal2018-003323
Application 14/735,002
Technology Center 2600
Before JOHNNY A. KUMAR, JENNIFER L. McKEOWN, and
CATHERINE SHIANG, Administrative Patent Judges.
McKEOWN, Administrative Patent Judge.
DECISION ON APPEAL
Appellants 1 appeal under 35 U.S.C. § 134(a) from the Examiner's
decision to reject claims 11-20. We have jurisdiction under 35 U.S.C. § 6.
We reverse.
1 According to Appellants, the real party in interest is Baidu USA, LLC.
App. Br. 3.
Appeal2018-003323
Application 14/735,002
STATEMENT OF THE CASE
Appellants' disclosed and claimed invention "relates to systems and
methods for improving the transcription of speech into text." Spec. ,r 3.
More specifically, Appellants describe that the claimed invention is directed
to
state-of-the-art speech recognition systems developed using end-
to-end deep learning. In embodiments, the model architecture is
significantly simpler than traditional speech systems, which rely
on laboriously engineered processing pipelines; these traditional
systems also tend to perform poorly when used in noisy
environments. In contrast, embodiments of the system do not
need hand-designed components to model background noise,
reverberation, or speaker variation, but instead directly learn a
function that is robust to such effects. A phoneme dictionary, nor
even the concept of a "phoneme," is needed. Embodiments
include a well-optimized recurrent neural network (RNN)
training system that can use multiple GPU s, as well as a set of
novel data synthesis techniques that allows for a large amount of
varied data for training to be efficiently obtained. Embodiments
of the system can also handle challenging noisy environments
better than widely used, state-of-the-art commercial speech
systems.
Abstract.
Claim 11 is illustrative of the claimed invention and reads as follows:
11. A computer-implemented method for transcribing
speech comprising:
receiving an input audio from a user;
normalizing the input audio to make a total power of the
input audio consistent with a set of training samples used to train
a trained neural network model;
generating a jitter set of audio files from the normalized
input audio by translating the normalized input audio by one or
more time values;
for each audio file from the jitter set of audio files, which
includes the normalized input audio:
2
Appeal2018-003323
Application 14/735,002
generating a set of spectrogram frames for each audio file;
inputting the audio file along with a context of
spectrogram frames into a trained neural network;
obtaining predicted character probabilities outputs from
the trained neural network; and
decoding a transcription of the input audio using the
predicted character probabilities outputs from the trained neural
network constrained by a language model that interprets a string
of characters from the predicted character probabilities outputs
as a word or words.
THE REJECTIONS
The Examiner rejected claims 11-20 under 35 U.S.C. § 101 as
directed to patent ineligible subject matter. Final Act. 6-14.
The Examiner rejected claims 11-20 under 35 U.S.C. § 103 as
unpatentable over Sompolinsky (US 2011/0035215 Al, published Feb. 10,
2011) and Talwar (2011/0282663 Al, published Nov. 17, 2011). Final Act.
7-14.
ANALYSIS
THE 35 U.S.C. § 101 REJECTION
Claims 11-20
Based on the record before us, we are persuaded that the Examiner
erred in rejecting claims 11-20 as directed to patent ineligible subject
matter.
An invention is patent-eligible if it claims a "new and useful process,
machine, manufacture, or composition of matter." 35 U.S.C. § 101.
However, the Supreme Court has long interpreted 35 U.S.C. § 101 to include
implicit exceptions: "[l]aws of nature, natural phenomena, and abstract
ideas" are not patentable. Alice Corp. v. CLS Bank Int 'l, 573 U.S. 208, 216
(2014).
3
Appeal2018-003323
Application 14/735,002
In determining whether a claim falls within an excluded category, we
are guided by the Supreme Court's two-step framework, described in Mayo
and Alice. Id. at 217-18 ( citing Mayo Collaborative Servs. v. Prometheus
Labs., Inc., 566 U.S. 66, 75-77 (2012)). In accordance with that framework,
we first determine what concept the claim is "directed to." See Alice, 573
U.S. at 219 ("On their face, the claims before us are drawn to the concept of
intermediated settlement, i.e., the use of a third party to mitigate settlement
risk."); see also Bilski v. Kappas, 561 U.S. 593, 611 (2010) ("Claims 1 and 4
in petitioners' application explain the basic concept of hedging, or protecting
against risk.").
Concepts determined to be abstract ideas, and thus patent ineligible,
include certain methods of organizing human activity, such as fundamental
economic practices (Alice, 573 U.S. at 219-20; Bilski, 561 U.S. at 611);
mathematical formulas (Parker v. Flook, 437 U.S. 584, 594--95 (1978)); and
mental processes (Gottschalkv. Benson, 409 U.S. 63, 69 (1972)). Concepts
determined to be patent eligible include physical and chemical processes,
such as "molding rubber products" (Diamond v. Diehr, 450 U.S. 175, 192
( 1981) ); "tanning, dyeing, making waterproof cloth, vulcanizing India
rubber, smelting ores" (id. at 184 n.7 (quoting Corning v. Burden, 56 U.S.
252, 267---68 (1854))); and manufacturing flour (Benson, 409 U.S. at 69
(citing Cochrane v. Deener, 94 U.S. 780, 785 (1876))).
In Diehr, the claim at issue recited a mathematical formula, but the
Supreme Court held that "[a] claim drawn to subject matter otherwise
statutory does not become nonstatutory simply because it uses a
mathematical formula." Diehr, 450 U.S. at 176; see also id. at 192 ("We
view respondents' claims as nothing more than a process for molding rubber
4
Appeal2018-003323
Application 14/735,002
products and not as an attempt to patent a mathematical formula."). Having
said that, the Supreme Court also indicated that a claim "seeking patent
protection for that formula in the abstract ... is not accorded the protection
of our patent laws, ... and this principle cannot be circumvented by
attempting to limit the use of the formula to a particular technological
environment." Id. (citing Benson and Flook); see, e.g., id. at 187 ("It is now
commonplace that an application of a law of nature or mathematical formula
to a known structure or process may well be deserving of patent
protection.").
If the claim is "directed to" an abstract idea, we tum to the second
step of the Alice and Mayo framework, where "we must examine the
elements of the claim to determine whether it contains an 'inventive
concept' sufficient to 'transform' the claimed abstract idea into a patent-
eligible application." Alice, 573 U.S. at 221 ( citation omitted). "A claim
that recites an abstract idea must include 'additional features' to ensure 'that
the [claim] is more than a drafting effort designed to monopolize the
[abstract idea]."' Id. ( quoting Mayo, 566 U.S. at 77). "[M]erely requir[ing]
generic computer implementation[] fail[ s] to transform that abstract idea into
a patent-eligible invention." Id.
The PTO recently published revised guidance on the application of
section 101. USPTO's January 7, 2019 Memorandum, 2019 Revised Patent
Subject Matter Eligibility Guidance, 84 Fed. Reg. 50 (Jan. 7, 2019)
("Memorandum"). Under that guidance, we first look to whether the claim
recites:
(1) any judicial exceptions, including certain groupings of abstract
ideas (i.e., mathematical concepts, certain methods of organizing
5
Appeal2018-003323
Application 14/735,002
human activity such as a fundamental economic practice, or mental
processes); and
(2) additional elements that integrate the judicial exception into a
practical application (see MANUAL OF PATENT EXAMINING PROCEDURE
(MPEP) § 2106.05(a)-(c), (e)-(h) (9th Ed., Rev. 08.2017, Jan. 2018)).
See Memorandum at 52, 55-56. Only if a claim (1) recites a judicial
exception and (2) does not integrate that exception into a practical
application, do we then look to whether the claim:
(3) adds a specific limitation beyond the judicial exception that are not
"well-understood, routine, conventional" in the field (see MPEP
§ 2106.05(d)); or
(4) simply appends well-understood, routine, conventional activities
previously known to the industry, specified at a high level of
generality, to the judicial exception.
See Memorandum at 56.
Examiner's Findings and Conclusion
Under step one of the Alice test, the Examiner determines that the
claims are directed to the abstract idea. Final Act. 2-3, 6-7. For example,
the Examiner identifies that the "claimed invention is directed to using the
predicted character probabilities (mathematical formula) to decode a
transcription of the input audio into words or text data." Final Act. 3; see
also Final Act. 6-7 (listing the limitations of claim 1 and finding that the
abstract idea "is similar to the court case Gottschalk v. Benson because the
predicted character probabilities (mathematical formula or relationship) is
used to convert or transcribe audio data into text data (words)."). The
Examiner determines that the Specification shows that the predicted
6
Appeal2018-003323
Application 14/735,002
character probabilities is an algorithm and, as such, the claimed invention is
directed to a mathematical formula. Final Act. 4 ( citing Spec. ,r,r 44, 93).
The Examiner also summarizes the claimed invention into three steps,
namely (1) "normalizing the input audio data (manipulating data)", (2)
"generating spectrogram frames based on each audio file (generating
information sets based on prior information sets)" and (3) "using a
mathematical formula to convert audio data into text data (Decoding)."
Ans. 4. According to the Examiner, "[ m ]anipulating data, generating
information based on prior information set and Decoding audio data using
equations or mathematical formula are all plainly abstract idea category of
judicial excepted subject matter" and the abstract ideas are categorized under
"'Certain Methods of Organizing Human Activity' since human can listen to
an audio file and transcribe the audio data into text data which can all be
done mentally." Ans. 4.
Under step two of the Alice test, the Examiner determines that the
claims do not amount to significantly more than the abstract idea. Final Act.
7. Specifically, the Examiner concludes that the claims "[ d]oes not amount
to significantly more since it is just decoding a transcription using a
mathematical formula or relationship (Predicted Character probabilities).
Thus converting or translating audio data into another form of data (text
data)." Final Act. 4.
Appellants ' Contentions
Appellants, on the other hand, maintain that the Examiner
overgeneralizes and oversimplifies the claimed invention and that the
claimed invention is not "directed to" an abstract idea. App. Br. 8-9. For
example, Appellants assert that the Examiner "tries to eliminate the trained
7
Appeal2018-003323
Application 14/735,002
neural network and related elements" by equating it to a generic computer.
According to Appellants,
A generic computer is not a trained neural network; but even
more, a generic computer is not the claimed trained neural
network that has been specially designed and trained to receive
sets of context of spectrogram frames from a jitter set of audio
files, which includes a normalized input audio file obtained from
an input audio, to predict character probabilities from the input
audio, which are finally selected by being constrained by a
language model that interprets a string of characters from the
predicted character probabilities outputs as a word or words.
Reply Br. 2.
Appellants also argue that the claimed invention is a specific
implementation to address specific technological problems in automatic
speech recognition (App. Br. 10-11) and is directed to a "specific
improvement in computer capabilities." App. Br. 14; see also Reply Br. 2--4
(citing the abstract, paragraphs 93-113, and Table 3 to support that the
claimed invention is "an improvement to the technical field of automatic
speech recognition."); App. Br. 11-14 (identifying that the claims here, like
the claims in Enfish and McRO, improve the functioning of prior art
automatic speech recognition systems).
With respect to step 2 of the Alice test, Appellants argue that the
claims include significantly more than the abstract idea. Appellants assert
that the claims "are not generic processes, such as merely storing and
retrieving data," but rather "the claims include specific implementations
steps" including the claimed normalizing, generating a jitter set, generating a
set of spectrogram frames, obtaining character probabilities, and decoding
using the predicted character probabilities. App. Br. 15. Moreover,
Appellants assert that the claims include significantly more because
8
Appeal2018-003323
Application 14/735,002
The claims of the current application deal with improvements to
a technology or technical field ( e.g., computer functionality via
automatic speech recognition, which may be used for numerous
purposes including computer interfacing); deal with
improvements to the functioning of the computer itself ( e.g., by
making it easier for users to interface with complex computing
systems); deal with a particular machine or system (e.g.,
computing systems that comprise specifically made models and
data); and add specific limitations other than what is well-
understood, routine and conventional in the field ( e.g., using a
specific and non-generic steps to obtain predicted character
probabilities outputs and decoding a transcription of the input
audio using the predicted character probabilities outputs from the
trained neural network constrained by a language model that
interprets a string of characters from the predicted character
probabilities outputs as a word or words).
App. Br. 16.
Analysis - Revised Step 2A
Under the Memorandum, in prong one of step 2A we look to whether
the claim recites a judicial exception. The Examiner identifies the abstract
ideas - a mathematical relationship/formula (Final Act. 3) and certain
methods of organizing human activity "since human can listen to an audio
file and transcribe the audio data into text data which can all be done
mentally." Ans. 4.
As an initial matter, we note that the Memorandum identifies mental
processes as a separate category of abstract ideas from methods of
organizing human activity. We disagree with the Examiner that the claims
recite either a method of organizing human activity or a mental process.
While transcription generally can be performed by a human, the claims here
are directed to a specific implementation including the steps of normalizing
an input file, generating a jitter set of audio files, generating a set of
9
Appeal2018-003323
Application 14/735,002
spectrogram frames, obtaining predicted character probabilities from a
trained neural network and decoding a transcription of the input audio using
the predicted character probability outputs. These are not steps that can
practically be performed mentally. Nor do we see how the claimed
invention recites organizing human activity. For example, the claims do not
include fundamental economic principles or practices, commercial or legal
interactions, managing personal behavior or relationships or interactions
between people. As such, the claims do not recite a mental process or
method of organizing human activity.
The claims do recite using predicted character probabilities to decide a
transcription of the input audio, which the Examiner, relying on the
Specification, determines is using a mathematical formula. Namely, the
Examiner identifies that the Specification discloses an algorithm to obtain
the predicted character probabilities. Final Act. 3--4 ( citing Spec. 44). The
mathematical algorithm or formula, however, is not recited in the claims. As
such, under the recent Memorandum, the claims do not recite a mathematical
concept. See, e.g., Subject Matter Eligibility Examples: Abstract Ideas, at 7
(Jan. 7, 2019)(discussing Example 38 and noting that "The claim does not
recite a mathematical relationship, formula, or calculation. While some of
the limitations may be based on mathematical concepts, the mathematical
concepts are not recited in the claims.").
Moreover, even if the claims were considered to recite a mathematical
concept, under prong two of step 2A the claims are not directed to an
abstract idea because the alleged judicial exception is integrated into a
practical application. Namely, as Appellants explain, "the claims of the
current application include specific features that were specifically designed
10
Appeal2018-003323
Application 14/735,002
to achieve an improved technological result" and "provide improvements to
that technical field." App. Br. 16. For example, the Specification describes
that using DeepSpeech learning, i.e. a trained neural network, along with a
language model "achieves higher performance than traditional methods on
hard speech recognition tasks while also being much simpler." Spec. ,r 29.
As such, based on the record before us, we are persuaded that the Examiner
erred in determining that the claims are directed to an abstract idea.
Analysis - Step 2B
We also agree with Appellants that the Examiner fails to sufficiently
support the finding that the claims do not add significantly more to the
alleged judicial exception. Namely, the Examiner concludes the claims do
not include "any additional elements that amounts to significantly more than
a judicial exception" but fails to provide sufficient factual support. See, e.g.,
Final Act. 4, 7; see also Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir.
2018).
Accordingly, we reverse the Examiner's decision to reject claims 11-
20 as directed to patent ineligible subject matter.
THE 35 U.S.C. § 103 OBVIOUSNESS REJECTION BASED ON SOMPOLINSKY AND
TALWAR
Based on the record before us, we are persuaded that the Examiner
erred in rejecting the claims as unpatentable over Sompolinsky and Talwar.
Appellants argue that Sompolinsky fails to teach or suggest decoding
a transcription of the input audio using the predicted character probabilities
outputs from the trained neural network constrained by a language model
that interprets a string of characters from the predicted character
11
Appeal2018-003323
Application 14/735,002
probabilities outputs as a word or words, as required by claims 11 and 17.
App. Br. 17. According to Appellants, "[t]he current application employs a
model that outputs character-level predictions" whereas Sompolinsky, in
contrast, "operates at a word-level or phoneme-level." App. Br 17.
We agree. Sompolinsky describes a speech recognition system that
includes neuron models where each neuron model is correlated with a pulse
pattern for a phoneme or entire word utterance. Sompolinsky i-fi-f l 9-20. In
other words, each phoneme or word will have a unique pulse pattern that
will be used to identify the phoneme or word. As such, Sompolinsky's
system outputs a phoneme, or a phonetic sound, not a character probability.
App. Br. 18. We are persuaded of error in the Examiner's determination that
Sompolinsky' s phoneme output would satisfy the claimed character level
probabilities.
We also agree with Appellants that the Examiner erred in relying on
Sompolinsky' s disclosure of "digits." See, e.g., Ans. 10 ( asserting that
Sompolinsky discloses transcription models based on digits and digits can be
characters). Sompolinsky merely discloses use of the numerical digit words,
such as "one" or "eight," as exemplary words for the neuron models. In
other words, it is the phoneme output for word, eight, not a digit
representation. The Examiner does not sufficiently identify any teaching of
the cited combination of Sompolinsky and Talwar that discloses the claimed
character level probabilities.
Accordingly, we reverse the Examiner's decision to reject claims 11-
20 as unpatentable over Sompolinsky and Talwar.
12
Appeal2018-003323
Application 14/735,002
DECISION
We reverse the Examiner's decision to reject claims 11-20 as directed
to patent ineligible subject matter and reverse the Examiner's decision to
reject claims 11-20 as unpatentable over Sompolinsky and Talwar.
REVERSED
13