JP3908878B2 - Phoneme recognition performance measuring device for continuous speech recognition device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a phoneme recognition performance measuring device for a continuous speech recognition device that recognizes continuous speech using an acoustic model and a language model.
[0002]
[Prior art]
Conventionally, a speech recognition apparatus that recognizes continuous speech using an acoustic model and a probabilistic language model is known (“speech recognition by a stochastic model”, Seiichi Nakagawa, IEICE). A functional configuration of this type of conventional speech recognition apparatus is shown in FIG.
[0003]
In FIG. 1, reference numeral 10 denotes an acoustic model that recognizes an input speech as a phoneme and outputs the recognized phoneme (label form) and its likelihood (likelihood). For example, HMM (Hidden Markov Model) is famous. . Reference numeral 11 denotes a language model. The language model has connection probabilities with words in the form of phoneme strings and adjacent words in the form of a database. 12 is the likelihood of the phoneme string input using the sequence and likelihood of the phonemes output from the acoustic model 10 in time series, the words described in the database of the language model 11 and their connection probabilities. It is an acoustic likelihood calculation unit that outputs the highest word in the language model as a word recognition result.
[0004]
In the continuous speech recognition apparatus having such a configuration, after continuous speech is phoneme-recognized in the acoustic model 10, the acoustic likelihood calculation unit 12 recognizes words and outputs a word string, that is, a sentence.
[0005]
Measuring the performance of acoustic models used in continuous speech recognizers is Matsuoka, Otsuki, Mori, Furui, Shirai, The Institute of Electronics, Information and Communication Engineers, December 2125, Matsuoka, Otsuki, Mori, Furui, Proposed in Shirai, Acoustical Society of Japan, March 1997, pages 2-6-11.
[0006]
[Problems to be solved by the invention]
To measure the performance of the acoustic model, input the speech for evaluation whose phoneme notation is known in advance into the acoustic model, and compare the phoneme string obtained as a result of phoneme recognition with the notation of the speech for evaluation. By doing so, the phoneme recognition rate can be obtained. Here, the phoneme refers to a phoneme such as a vowel or a consonant, or a voice part having a shorter voice length than the phoneme. A phoneme output from the acoustic model is an identification name (label) of a phoneme recognized. ) Etc.
[0007]
The use of the phoneme recognition rate of the acoustic model as an index indicating the phoneme recognition performance of the continuous speech recognition apparatus has the following problems.
[0008]
In the acoustic model 10, phoneme recognition is performed by comparing an acoustic feature obtained from input speech with an acoustic feature of a phoneme that is held in advance in the acoustic model.
[0009]
Since the acoustic model alone performs only phoneme recognition and does not perform word recognition, a phoneme misrecognition result that cannot be a linguistic sentence or word connection occurs.
[0010]
In the continuous speech recognition apparatus, such a misrecognition result is corrected in the word recognition of the acoustic likelihood calculation apparatus 12. Therefore, when the phoneme recognition rate of the acoustic model alone is used as the phoneme recognition performance measurement value in the continuous speech recognition apparatus, the measurement value tends to be lower than the actual performance.
[0011]
In view of the above points, an object of the present invention is to provide a continuous speech recognition apparatus capable of correctly measuring the phoneme recognition performance of a continuous speech recognition apparatus when a grammatical connection relationship (word connection rate) in a language is excluded. and to provide a phoneme recognition performance measurement equipment.
[0012]
[Means for Solving the Problems]
In order to achieve such an object, in the phoneme recognition performance measuring device of a continuous speech recognition device that measures the phoneme recognition performance of a continuous speech recognition device that performs continuous speech recognition using a language model and an acoustic model, An input means for inputting a notation of a phoneme string, and the evaluation speech is input to the continuous speech recognition device, phoneme recognition is performed on the evaluation speech by the acoustic model, and the phoneme recognition of the acoustic model is performed. Phoneme processing means for constraining phonemes based on character strings existing in the result, phoneme strings output from the phoneme processing means, and phoneme strings of evaluation speech input from the input means the phoneme recognition rate calculated by the comparison and representation, the calculated phoneme recognition rate comprises a performance measuring means for outputting a measure of phoneme recognition performance, or the phoneme restrained by the phoneme processing means Excludes the phoneme sequence that does not exist on the language by Rukoto, characterized by the use of processing results of the phoneme processing means excluding phoneme sequence which does not exist on該言words the performance measurement of the performance measurement means.
[0013]
According to a second aspect of the present invention, in the phoneme recognition performance measuring apparatus of the continuous speech recognition apparatus according to the first aspect, the character string existing in the language is a word.
[0016]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0017]
FIG. 2 shows a basic configuration of a phoneme recognition performance measuring apparatus provided in the continuous speech recognition apparatus. 2, the same parts as those in the conventional example of FIG. 1 are denoted by the same reference numerals, and detailed description thereof is omitted. A phoneme string generation unit 21 having a word dictionary 21-1 sequentially supplies the phoneme string form words described in the word dictionary 21-1 to the acoustic likelihood calculation unit 12. As a result, the acoustic likelihood calculation unit 12 compares the plurality of phonemes output from the acoustic model 10 for a certain period of time, that is, the phoneme sequence and the phoneme sequence of the word supplied from the phoneme sequence generation unit 21, and is most similar. The word (phoneme string form) is output as a word recognition result.
[0018]
In this embodiment, since the connection rate of words is not given to the acoustic likelihood calculation unit 12, the connection relationship of words in the language is not considered, and the phoneme string of words that may exist in the language is obtained as a speech recognition result. Please note that. Therefore, for example, even if a phoneme “k” that does not exist in the language such as “aaaskssaa” is mixed as a phoneme recognition result of the acoustic model for the evaluation voice “morning”, If a phoneme sequence corresponding to “morning” is present in the word supplied from the “sequence generation unit” 21, the phoneme “k” in the phoneme recognition result is corrected, and “aaasssaaa” is output as the word recognition result. The In this specification, performing phoneme processing that excludes phonemes that do not exist in the language is referred to as “constraining the phoneme recognition result of the acoustic model”. In this embodiment, the length of the character string to be constrained is a word, but it may be a phrase or a sentence depending on the application. In this case, the word dictionary 21-1 used in the phoneme string generation unit 21 is compatible. It describes a phoneme string of character string length.
[0019]
A phoneme string comparison unit 22 compares the word recognition result in the phoneme string form output from the acoustic likelihood calculation unit 12 when the evaluation speech is input with the notation of the speech for evaluation, and performs phoneme recognition. Calculate the rate. As a comparison method, in this embodiment, dynamic programming (also called DP or dynamic programming) is used. The calculated recognition rate is visually output by a display or printer (not shown).
[0020]
An example of a hardware configuration for realizing a continuous speech recognition apparatus having such a phoneme recognition performance measuring apparatus is shown in FIG. The continuous speech recognition apparatus is realized by mounting a continuous speech recognition program on a general-purpose computer such as a personal computer. FIG. 3 shows a configuration of main parts of a personal computer or the like.
[0021]
In FIG. 3, the CPU 100 executes a speech recognition program to perform continuous speech recognition processing, and also executes a program for phoneme recognition performance described later to measure the phoneme recognition performance of the continuous speech recognition apparatus.
[0022]
The system memory 110 temporarily stores input / output data for information processing performed by the CPU 100. A hard disk storage device (abbreviated as HDD) 130 is a word dictionary (corresponding to 21-1 in FIG. 2) and phonemes related to the present invention, such as a continuous speech recognition program, data for a language model, and data used in an acoustic model. Save and store the recognition performance measurement program. The above-described program is an instruction to execute a keyboard or mouse (not shown), and after being loaded from the HDD 130 into the system memory 110, the program instructed by the CPU 100 is executed.
[0023]
An input interface (I / O) 120 A / D converts an audio signal input from a microphone (not shown) and delivers the digital audio signal to the CPU 100 for continuous audio recognition.
[0024]
In such a configuration, when the user instructs execution of the phoneme recognition performance measurement program using a keyboard or a mouse, the phoneme recognition performance measurement program of FIG. 4 is loaded from the HDD 130 to the system memory 110 and executed by the CPU 100.
[0025]
Assume that the user stores in advance the phonetic notation of the voice for evaluation in the HDD 130 in the form of a document file. The phonetic document file may be created by a word processor software of a personal computer, or may be transferred off-line to the HDD 130 (not shown) via a floppy or the like.
[0026]
The user inputs evaluation voice from a microphone (not shown). The input voice is delivered to the CPU 100 via the I / O 120. The CPU 100 performs phoneme recognition using an acoustic model portion (for example, a form of a subroutine or function) in a continuous speech recognition program similar to the conventional one, and sequentially stores the phoneme recognition results in a work area in the system memory 110. I will do it.
[0027]
The CPU 100 extracts the word (phoneme string form) that is most similar to the phoneme recognition result from the word dictionary 21-1. This processing corresponds to the processing of the phoneme string generation unit 21 and the acoustic likelihood calculation unit 22 in FIG. More specifically, the phoneme strings of words are sequentially read out from the word dictionary 21-1, and compared with the output of the comparison acoustic model 10, that is, the phoneme string, and the likelihood is calculated in the same manner as in the prior art. To do. Thereafter, the words are sequentially read from the word dictionary, and compared with the phoneme string to be compared, the likelihood is calculated. The likelihood calculated first is temporarily stored as a temporary maximum value together with the phoneme string and / or the storage position of the word dictionary 21-1. Next, the likelihood for the read word is compared with the temporary maximum value, and if the likelihood read next time is larger than the temporary maximum value, the phoneme string of the read word and Alternatively, the storage position and likelihood of the word are replaced with the temporary maximum value and related data temporarily stored so far.
[0028]
In this way, all the phoneme strings of words described in the word dictionary 21-1 are compared with the comparison phoneme string, that is, the likelihood calculation is performed. Then, the provisional maximum value stored in the system memory 110, the corresponding phoneme string (notation), and the storage position of the word dictionary are determined as the word recognition result.
[0029]
The determined phoneme string is stored in the speech recognition result storage area of the system memory 110 (step S10 in FIG. 4).
[0030]
When the above processing is repeated at regular time intervals until the end of speech input (loop processing in steps S10 to S20), the word recognition results are accumulated in the form of phoneme strings in the speech recognition result storage area of the system memory 110. Will be memorized.
[0031]
Next, when the speech recognition for the input speech is completed, the CPU 100 reads (inputs) a phoneme-notation file of evaluation speech from the HDD 130 onto the work area of the system memory 110. Subsequently, the CPU 100 uses the phoneme string temporarily stored in the word recognition result storage area of the system memory 110 and the phoneme notation of the evaluation speech stored in the work area to perform dynamic programming, that is, DP The phoneme recognition rate is calculated by a matching method (steps S30 to S40 in FIG. 4).
[0032]
The calculated phoneme recognition rate is output as a measurement value of the phoneme recognition performance to a display or a printer (step S50 in FIG. 4).
[0033]
The user can know the phoneme recognition performance of the phoneme recognition function with the word constraint used in the continuous speech recognition program by looking at the output measurement values.
[0034]
In addition to this embodiment, the following embodiment can be implemented.
[0035]
1) In the above-described embodiment, the continuous speech recognition program and the phoneme recognition performance measurement program are separated, but a phoneme recognition performance measurement program can be incorporated into the continuous speech recognition program. In this case, a mode instruction indicating whether to perform continuous speech recognition using the keyboard or mouse or to perform phoneme recognition performance is received from the user. When continuous speech recognition is instructed, the acoustic likelihood calculation unit 12 of the continuous speech program is defined to use a language model, and when phoneme recognition performance measurement is instructed, switching to the phoneme string generation unit is performed. (See FIG. 5).
[0036]
Furthermore, in the above embodiment, the word connection probability output from the language model 11 is ignored, but a fixed value is given to the acoustic likelihood calculation unit 12 as the word connection probability when measuring the phoneme recognition performance as shown in FIG. The word information in the phoneme string form of the language model can be used instead of the word dictionary (21-1) used for measuring the phoneme recognition performance.
[0037]
2) In the above embodiment, the recording medium for recording the voice recognition performance measurement program may be a hard disk storage device (HDD), but the recording medium may be an IC memory such as a ROM or a RAM, a floppy disk or a CDROM, or the like. It may be a portable recording medium.
[0038]
3) In the above-described embodiment, the phoneme recognition rate is a measured value of the phoneme recognition performance, but the value of the phoneme recognition rate itself may be used, and the phoneme recognition rate may be, for example, good, normal, or bad. It may be expressed in stages.
[0039]
4) The above-described embodiment is an embodiment for explaining the present invention, and modifications other than the above can be made to the above-described embodiment in accordance with the technical idea indicated by the claims. Also, such variations are within the scope of the present patent.
[0040]
【The invention's effect】
As described above, according to the present invention, for example, the phoneme recognition result of the acoustic model is constrained in units of words (phonemes that do not exist in the language for character strings such as words are determined from the phoneme recognition result. This makes it possible to correctly measure the phoneme recognition performance of the continuous speech recognition apparatus.
[0041]
In addition, when analyzing the recognition performance of the continuous speech recognition device as a whole, the recognition performance of the continuous speech recognition device can be efficiently analyzed by referring to the measurement results of the phoneme recognition performance of the continuous speech recognition device. .
[Brief description of the drawings]
FIG. 1 is a block diagram showing a typical configuration of a conventional speech recognition apparatus.
FIG. 2 is a block diagram showing a basic configuration of the embodiment of the present invention.
FIG. 3 is a block diagram showing a hardware configuration of an embodiment of the present invention.
FIG. 4 is a flowchart showing a phoneme recognition performance measurement processing procedure of a CPU.
FIG. 5 is a block diagram showing another embodiment of the present invention.
FIG. 6 is a block diagram showing still another embodiment of the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 10 Acoustic model 11 Language model 12 Acoustic likelihood calculation part 21 Phoneme sequence generation part 21-1 Word dictionary 100 CPU
110 System memory 120 I / O
130 HDD

Claims (2)

In the phoneme recognition performance measuring device of a continuous speech recognition device that measures the phoneme recognition performance of a continuous speech recognition device that performs continuous speech recognition using a language model and an acoustic model,
An input means for inputting a phoneme string notation of the evaluation voice;
The evaluation speech is input to the continuous speech recognition apparatus, phoneme recognition is performed on the evaluation speech by the acoustic model, and a language string is present for a phoneme recognition result of the acoustic model. Phoneme processing means for constraining the phoneme
A phoneme recognition rate is calculated by comparing the phoneme sequence output from the phoneme processing unit with the notation of the phoneme sequence of the evaluation speech input from the input unit, and the calculated phoneme recognition rate is used as a phoneme recognition performance. A phoneme processing means that excludes phoneme strings that do not exist in the language and excludes phoneme strings that do not exist in the language by constraining the phonemes by the phoneme processing means. A phoneme recognition performance measuring apparatus for a continuous speech recognition apparatus , wherein a processing result is used for performance measurement of the performance measuring means.   The phoneme recognition performance measuring apparatus of the continuous speech recognition apparatus according to claim 1, wherein the character string existing in the language is a word.

Images