Question and answer database expansion based on speech recognition using a specialized and a general language model

• US 8,515,764 B2
• Filed: 07/08/2010
• Issued: 08/20/2013
• Est. Priority Date: 07/08/2009
• Status: Active Grant

• Current Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Original Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Inventors: Nakano, Mikio, Funakoshi, Kotaro, Narimatsu, Hiromi
• Primary Examiner(s): Smits, Talivaldis Ivars
• Application Number:
  US12/832,426
• Publication Number:
  US 20110010177A1
• Time in Patent Office: 1,139 Days
• Field of Search: None
• US Class Current: 704/275
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

A question and answer database expansion apparatus includes: a question and answer database in which questions and answers corresponding to the questions are registered in association with each other, a first speech recognition unit which carries out speech recognition for an input sound signal by using a language model based on the question and answer database, and outputs a first speech recognition result as the recognition result, a second speech recognition unit which carries out speech recognition for the input sound signal by using a language model based on a large vocabulary database, and outputs a second speech recognition result as the recognition result, and a question detection unit which detects an unregistered utterance, which is not registered in the question and answer database, from the input sound based on the first speech recognition result and the second speech recognition result, and outputs the detected unregistered utterance.

Claims

1. A question and answer database expansion apparatus comprising:

a question and answer database in which questions and answers corresponding to the questions are registered in association with each other;

a first speech recognition unit which carries out speech recognition for an input sound signal by using a language model based on the question and answer database, and outputs a first speech recognition result as the recognition result;

a second speech recognition unit which carries out speech recognition for the input sound signal by using a language model based on a large vocabulary database, and outputs a second speech recognition result as the recognition result; and

a first answer selection unit which determines whether or not a question based on the first speech recognition result and an answer corresponding to the question are registered in the question and answer database;

a second answer selection unit which determines whether or not a question based on the second speech recognition result and an answer corresponding to the question are registered in the question and answer database,

wherein the question detection unit calculates a first feature quantity based on the first speech recognition result, calculates a second feature quantity based on the second speech recognition result, and detects an utterance, which is not included in the question and answer database, by using the calculated first feature quantity, the calculated second feature quantity, the determination result of the first answer selection unit, and the determination result of the second answer selection unit; and

a question detection unit which detects an unregistered utterance, which is not registered in the question and answer database, from the input sound based on the first speech recognition result and the second speech recognition result, and outputs the detected unregistered utterance.

2. The question and answer database expansion apparatus according to claim 1,

wherein the question detection unit detects an unregistered utterance, which is not registered in the question and answer database, from the input sound in terms of sentences.

3. The question and answer database expansion apparatus according to claim 1,

wherein the question detection unit carries out learning by using the first feature quantity and the second feature quantity, calculates a determination function for determining whether being registered in the question and answer database or not based on the learned result, and detects an utterance, which is not included in the question and answer database, based on an evaluation value calculated by the determination function.

4. The question and answer database expansion apparatus according to claim 1,

wherein the first feature quantity is at least one of values obtained based on presence/absence of an utterance based on a supportive response obtained when the speech recognition is carried out, the length of the utterance, an acoustic score obtained when the speech recognition is carried out, a language score obtained when the speech recognition is carried out, recognition precision of the first speech recognition result, and confidence of the first speech recognition result, and the second feature quantity is at least one of values obtained based on presence/absence of a supportive response obtained when the speech recognition is carried out, the length of an utterance, an acoustic score obtained when the speech recognition is carried out, and a language score obtained when the speech recognition is carried out.

5. A question and answer database expansion method for a question and answer database expansion apparatus using a question and answer database, in which questions and answers corresponding to the questions are registered in association with each other, the question and answer database expansion method comprising:

a first speech recognition step of causing a first speech recognition unit to carry out speech recognition for an input sound signal by using a language model based on the question and answer database and to output a first speech recognition result as the recognition result;

a second speech recognition step of causing a second speech recognition unit to carry out speech recognition for the input sound signal by using a language model based on a large vocabulary database and to output a second speech recognition result as the recognition result a first answer selection step which determines whether or not a question based on the first speech recognition result and an answer corresponding to the question are registered in the question and answer database; and

a second answer selection step which determines whether or not a question based on the second speech recognition result and an answer corresponding to the question are registered in the question and answer database,

wherein the question detection step calculates a first feature quantity based on the first speech recognition result, calculates a second feature quantity based on the second speech recognition result, and detects an utterance, which is not included in the question and answer database, by using the calculated first feature quantity, the calculated second feature quantity, the determination result of the first answer selection unit, and the determination result of the second answer selection unit; and

a question detection step of causing a question detection unit to detect an unregistered utterance, which is not registered in the question and answer database, from the input sound based on the first speech recognition result and the second speech recognition result and to output the detected unregistered utterance.