System and method of lattice-based search for spoken utterance retrieval
AssigneeAT&T Intellectual Property II, L.P.
InventorsMurat Saraclar, Richard William Sproat
Life of this patent
Estimating the term…
Abstract
A system and method are disclosed for retrieving audio segments from a spoken document. The spoken document preferably is one having moderate word error rates such as telephone calls or teleconferences. The method comprises converting speech associated with a spoken document into a lattice representation and indexing the lattice representation of speech. These steps are performed typically off-line. Upon receiving a query from a user, the method further comprises searching the indexed lattice representation of speech and returning retrieved audio segments from the spoken document that match the user query.
Claims23
A method of retrieving a spoken document, the method being performed by a computing device, the method comprising:
- converting via a processor high word error rate speech from the spoken document into a lattice representation related to a recognition network represented as weighted finite state machines, the spoken document generated from a telephone call;
- indexing via the processor the lattice representation of speech based on phones;
- upon receiving a query from a user:
- converting each query word into phone strings based on the query word pronunciation;
- searching the phone-based index of the lattice representation of speech for each phone string of the phone strings having a minimum pronunciation length;
- searching the indexed lattice representation of speech;
- and
- returning audio segments from the spoken document that match the user query.
A method of retrieving a spoken document wherein a word index and a sub-word index related to the spoken document exist, and are related to a recognition network represented as weighted finite state machines, wherein the word index and sub-word index are generated based on high word error rate speech from the spoken document, the method comprising, upon receiving a query from a user:
- converting via a processor each query word into phone strings based on the query word pronunciation, the spoken document generated from a telephone call;
- searching via the processor the phone-based index of lattice representation of speech for each phone string of the phone string having a minimum pronunciation length;
- searching the word index based on the user query;
- searching the sub-word index based on the user query;
- and
- combining the results to retrieve the audio segments from the spoken document that match the user query.
A method of retrieving a spoken document wherein a word index and a sub-word index related to the spoken document exist and are related to a recognition network represented as weighted finite state machines, the method comprising, upon receiving a user query from a user:
- converting via a processor each query word from the user into phone strings based on a query word pronunciation, the spoken document generated from a telephone call;
- searching the word index based on the user query if the user query is in-vocabulary;
- searching the sub-word index based on the user query if the user query for each phone string of the phone string having a minimum pronunciation length is out of vocabulary;
- and
- retrieving the spoken document based on one of the searched word index on the searched sub-word index.
A system for retrieving a spoken document, the system comprising:
- a processor;
- a first module controlling the processor to convert high word error rate speech from the spoken document into a lattice representation related to a recognition network represented as weighted finite state machines, the spoken document generated from a telephone call;
- a second module controlling the processor to index the lattice representation of speech;
- upon receiving a query from a user:
- a third module controlling the processor to convert each query word into phone strings based on the query word pronunciation;
- a fourth module controlling the processor to search the phone-based index of lattice representation of speech for each phone string of the phone strings having a minimum pronunciation length;
- a fifth module controlling the processor to search the indexed lattice representation of speech;
- and
- a sixth module controlling the processor to return audio segments from the spoken document that match the user query.
A computer-readable storage medium storing a computer program having instructions for controlling a computing device to retrieve a spoken document, the instructions causing the computing device to perform a method comprising:
- converting high word error rate speech from the spoken document into a lattice representation related to a recognition network represented as weighted finite state machines, the spoken document generated from a telephone call;
- indexing the lattice representation of speech;
- upon receiving a query from a user:
- converting each query word into phone strings based on the query word pronunciation;
- searching the phone-based index of lattice representation of speech for each phone string of the phone strings having a minimum pronunciation length;
- searching the indexed lattice representation of speech;
- and
- returning audio segments from the spoken document that match the user query.
Dependencies are inferred from the claim text.