JPH05188998A - Speech recognizing method - Google Patents

Abstract

PURPOSE:To shorten the matching process time by employing a neural network for a matching process between a phoneme series obtained from an acoustic process part and the phoneme series of a dictionary. CONSTITUTION:The phoneme series sent out of the acoustic process part which extracts features from an input speech and converts them into the phoneme series by segmentation and phoneme recognition is supplied to a phoneme series input part 13 consisting of the neural network 12 which performs the matching process. The neural network 12 learns the phoneme series from the dictionary input part 14 which contains the previously registered phoneme series and the phoneme series input part 13 and obtains the proper phoneme series from the phoneme series from the phoneme series input part 13 which includes errors at an output part 15. Namely, the phoneme series obtained from the acoustic process part 11 is matched by the neural network 12 with the phoneme series from the dictionary input part 14. Then the phoneme series obtained at the output part 15 is supplied to a language process part 16, which generates a correct document by using language information on syntax, meaning, context, etc.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous speech recognition method, and more particularly to a speech recognition method having an improved phoneme sequence matching process.

[0002]

2. Description of the Related Art A general structure of a continuous speech recognition method is a speech processing unit for obtaining a phoneme sequence from input speech data, and a language processing for constructing a word from the phoneme sequence obtained by the sound processing unit to generate a sentence. Consists of parts. The language processing unit cuts out words and compares the phoneme sequence input from the obtained sound processing unit with the phoneme sequence registered in the dictionary to generate a sentence. At that time, since the phoneme sequence transmitted from the acoustic processing unit may include an error, the phoneme sequence is processed while performing a brute force match with the phoneme sequence registered in the dictionary.

[0003]

Since the phoneme sequence sent from the acoustic processing unit and the phoneme sequence registered in the dictionary are subjected to the brute force matching process as described above, the processing time is extremely long. There is a problem that becomes.

The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a speech recognition method designed to shorten the phoneme sequence matching processing time.

[0005]

In order to achieve the above object, the present invention converts an input speech into a phoneme sequence in a sound processing unit and then supplies this phoneme sequence to one input unit of a neural network. Then, the other input unit is supplied with the phoneme sequence from the dictionary that stores the registered phoneme sequences, and the matching process of both phoneme sequences is performed by the neural network, and then the output unit of the neural network is set to the language processing unit. It is characterized in that it is supplied to generate a sentence.

[0006]

The phoneme sequence sent from the acoustic processing unit is given to one phoneme sequence input unit of the neural network, and the registered phoneme sequence is given to the other dictionary input unit. The neural network learns both phoneme sequences to obtain an appropriate phoneme sequence at the output section. The language processing unit generates a sentence based on this phoneme sequence.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. In FIG. 1, reference numeral 11 denotes an acoustic processing unit that performs feature extraction from input speech and converts it into a phoneme sequence by segmentation and phoneme recognition. The neural network 1 that performs matching processing on the phoneme sequence sent from this acoustic processing unit 11
It is given to the phoneme sequence input unit 13 consisting of two. Reference numeral 14 is a dictionary input unit that also stores a pre-registered phoneme sequence, which is also a neural network, and the neural network 12 learns the phoneme sequences from both input units 13 and 14 to eliminate errors from the phoneme sequence input unit 13. An appropriate phoneme sequence is obtained from the included phoneme sequence at the output unit 15. The phoneme sequence obtained at the output unit 15 is supplied to the language processing unit 16, where a correct sentence is generated using language information such as syntax, meaning, and context.

Next, in the embodiment configured as described above, since the phoneme sequence from the acoustic processing unit 11 contains an error, this phoneme sequence is learned by the neural network 12 by the phoneme sequence of the dictionary input unit 14. For example, it is output to the output unit 15 of the neural network 12 so that the part where the phoneme sequence matches becomes “1” and the other parts become “0”. The language processing unit 16 generates a sentence using the appropriate phoneme sequence obtained by the output unit 15. In this way, if both phoneme sequences are subjected to matching processing by the neural network 12, the processing time will be greatly shortened.

[0009]

As described above, according to the present invention,
Since the neural network is used for matching processing between the phoneme sequence obtained from the acoustic processing unit and the phoneme sequence in the dictionary, the matching processing time can be shortened and the matching applied to the error tendency in the acoustic processing unit can be achieved. There are advantages such as realization.

[Brief description of drawings]

FIG. 1 is a structural explanatory view showing an embodiment of the present invention.

[Explanation of symbols]

11 ... Acoustic processing unit, 12 ... Neural network, 1
3 ... Phoneme sequence input unit, 14 ... Dictionary input unit, 15 ... Output unit, 16 ... Language processing unit.

Claims (1)

[Claims] 1. A sound processing unit converts an input speech into a phoneme sequence, and then supplies this phoneme sequence to one input unit of a neural network, and stores the registered phoneme sequence in the other input unit. A speech recognition method comprising supplying a phoneme sequence from the dictionary and performing matching processing of both phoneme sequences by a neural network, and then supplying an output unit of the neural network to a language processing unit to generate a sentence.