The following operators can be used to better focus your queries.
( ) , AND, OR, NOT, W/#
? single char wildcard, not at start
* multi char wildcard, not at start
(Cat? OR feline) AND NOT dog?
Cat? W/5 behavior
(Cat? OR feline) AND traits
Cat AND charact*
This guide provides a more detailed description of the syntax that is supported along with examples.
This search box also supports the look-up of an IP.com Digital Signature (also referred to as Fingerprint); enter the 72-, 48-, or 32-character code to retrieve details of the associated file or submission.
Concept Search - What can I type?
For a concept search, you can enter phrases, sentences, or full paragraphs in English. For example, copy and paste the abstract of a patent application or paragraphs from an article.
Concept search eliminates the need for complex Boolean syntax to inform retrieval. Our Semantic Gist engine uses advanced cognitive semantic analysis to extract the meaning of data. This reduces the chances of missing valuable information, that may result from traditional keyword searching.
A technique is described which employs minimum training data to obtain normalizing transformation, thereby permitting new talkers to operate speech recognition systems.
English (United States)
This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately
55% of the total text.
Page 1 of 3
Fast Training Method For Speech Recognition Systems
A technique is described which employs minimum training data to obtain
normalizing transformation, thereby permitting new talkers to operate speech
There exists a real need for a normalizing transformation which would make
a new talker appear like a known talker to the system or, equivalently, to make
the system's reference patterns appear like those of the new talker. For such
transformations to be viable, they must yield reasonable accuracies for new
talkers and require a relatively small amount of training data. The normalization
algorithm operates on raw acoustic measurements normally extracted from the
speech signal approximately every 10 msec. The small training set consists of
either a short paragraph or a small number of discrete utterances.
A block diagram of the steps of the procedure is shown in the
figure. The training data from a new talker is time aligned per step
10 with that of a talker known to the system by means of a standard
dynamic programming procedure 12 such as that described in (*).
Given this alignment, the least squares approximation technique 14 is
used to find the transformation which minimizes in the least square
sense the difference between the training data for the new talker and
the transformed training data for the known talker. Specifically, we
want to transform each measurement vector from talker A into one
approximating talker B, that is
where a is an
n-dimensional measurement vector from talker A,
n-dimensional estimated measurement vector representing talker B, and
T is an nxn matrix. T is obtained from a limited set of training
data from talkers A a...