Browse Prior Art Database

Radio System with Coded Data and Speech-Recognition

IP.com Disclosure Number: IPCOM000115977D
Original Publication Date: 1995-Jul-01
Included in the Prior Art Database: 2005-Mar-30
Document File: 2 page(s) / 100K

Publishing Venue

IBM

Related People

Cohen, PS: AUTHOR [+2]

Abstract

Disclosed is a radio system in which digitally coded data, transmitted along with radio broadcast signals, is used to transmit phonological labels for the names of stations, shows, people, topics, etc., to transmit actual speaker-independent acoustic models or proper names and words, and to flag and identify words in a broadcast. An example of the transmission of coded data is found in the transmissions of the Radio Broadcast Data System (RBDS). A radio receiver in this system uses a look-up table or translation algorithm to decode the coded data and to associate phonological data with the coded data.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 51% of the total text.

Radio System with Coded Data and Speech-Recognition

      Disclosed is a radio system in which digitally coded data,
transmitted along with radio broadcast signals, is used to transmit
phonological labels for the names of stations, shows, people, topics,
etc., to transmit actual speaker-independent acoustic models or
proper names and words, and to flag and identify words in a
broadcast.  An example of the transmission of coded data is found in
the transmissions of the Radio Broadcast Data System (RBDS).  A radio
receiver in this system uses a look-up table or translation algorithm
to decode the coded data and to associate phonological data with the
coded data.

      Many speech recognition systems use an internal coding
technique to represent the phonological sounds in a language.  For
example, "TSH" may represent the "ch" sound in "children."  While
these internal coding techniques slightly expand the number of bits
required to describe a word, they are generally efficient ways of
transmitting or describing the acoustics and vectors associated with
proper names.  In the presently described radio system, these
phonological descriptors are transmitted as coded data.  Within the
radio receiver, this phonological data is then imbedded as a variable
name in Backus-Naur Form (BNF) grammars, which may subsequently used
in responding to commands, such as, "Find <variable name>," or, "Tune
to channel <variable name>."

      Furthermore, different categories of variable names can be
imbedded in different BNFs and in different contexts to limit
perplexity, which is a function of the number of active, competing
words (1,2).  For example, in the command, "Listen for the <variable
name> score," the perplexity of the variable name is reduced because
the word "score" at the end limits the active vocabulary of the
variable name to the names of active sports teams.

      Some speech recognition systems, instead of using an internal
phonetic labeling technique, keep the actual acoustic or waveform
data, subsequently attempting to perform a "best match" against it.
Some systems ask the user to repeat words or phrases several times,
capturing the acoustics of these samples to perform a "match-back"
against these waveforms, using Hidden Markov Modeling techniques (3).
This method can be quite effective when a large sample of individuals
say a proper name, so that a "golden voice" for that word or phrase
can be created.  In the presently-disclosed radio system, this
"golden voice" technique can be used instead of the transmission of
phonetic labels.  Once this is done, the spotting of proper names
proceeds as described above.  While this method uses more b...