Title:
SPEECH RECOGNITION DEVICE
Document Type and Number:
Japanese Patent JP3231365
Kind Code:
B2
Abstract:
PURPOSE: To obtain a high recognition rate by a learning function (word learning) wherein the time segmentation of an input into phonemes by phoneme dictionary learning and top-down recognition is considered in speech recognition based upon phonemes as units.
CONSTITUTION: The speech recognition device consists of means 1 and 2 which extract phoneme feature vectors from a time series of feature parameters of a speech found by analyzing input speech data, means 3 and 5 which collate the extracted phoneme feature vectors with a phoneme dictionary to obtain a likelihood time series of phonemes, means 4 and 6 which collate the phoneme likelihood time series with a word dictionary to output the recognition output and obtain phoneme section information on the input from the result of collation between the phoneme likelihood time series and word dictionary, a means 13 which selects and extracts phoneme feature vectors for learning from the phoneme section information and the time series of the feature parameters of the input, and means 7, 8, and 11 which collates the phoneme feature vectors for learning with the phoneme dictionary and learns the phoneme dictionary with the decision result of the phoneme collation and the phoneme feature vectors for learning to obtain a new phoneme dictionary.
Inventors:
Hiroyuki Tsuboi
Application Number:
JP25255591A
Publication Date:
November 19, 2001
Filing Date:
September 30, 1991
Export Citation:
Assignee:
Toshiba Corporation
International Classes:
G10L15/06; G10L15/10; G10L15/18; G10L15/183; (IPC1-7): G10L15/18; G10L15/06
Domestic Patent References:
| JP6337396A | ||||
| JP61133996A | ||||
| JP62278597A |
Other References:
【文献】古井「ディジタル音声処理」東海大学出版会(昭和60年9月)第167~169頁
Attorney, Agent or Firm:
Takehiko Suzue