JP2712586B2 - Pattern matching method for word speech recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pattern registration type word speech recognition apparatus,
In particular, the present invention relates to a pattern matching method in which a word standard pattern is divided into a consonant part, a transient part, and a vowel stationary part, and matching is performed linearly or nonlinearly adaptively.

[Prior Art] Conventionally, this type of word voice confirmation device has a configuration as shown in FIG. In the figure, 1 is a voice input unit such as a microphone, 2 is a start / end detection unit that separates an input voice signal from a silent section, 3 is an acoustic analysis unit that extracts feature parameters suitable for recognition from the voice signal, 10 Is a pattern matching unit that performs DP matching etc. between the standard pattern and the extracted feature parameters, 6 is a word standard pattern memory unit that stores standard patterns to be registered, and 8 is a recognition unit that obtains recognition results based on the results of pattern matching. It is a result processing unit.

[Problems to be Solved by the Invention] In the conventional word speech recognition apparatus described above, a non-linear matching method represented by DP matching is generally adopted as a pattern matching method. Non-linear matching methods such as DP matching are highly adaptable to the expansion and contraction of the pattern in the time axis direction due to the difference in the generation speed, and can obtain a high word recognition rate, compared to the linear matching method There is an advantage.

However, the expansion and contraction of the word pattern due to the difference in the utterance speed almost entirely occurs in the stationary vowel part, and the consonant and the transition part do not undergo much temporal expansion and contraction. For this reason, when performing non-linear matching over the entire word pattern, the weight of the vowel portion pattern becomes large, and the ratio of the consonant and the transient portion to the evaluation value is relatively reduced, which may cause erroneous recognition. For example, / gosjaku /
(5 shaku) and / osjaku / (ladle), / jonnijon / (424) and / jon
Words with similar vowel sequences, such as jon / (44), are prone to misrecognition. In order to avoid this, DP matching takes measures such as setting a slope limit on the matching path, but there is no definitive solution at present.

[Means for Solving the Problems] A pattern matching method for a word speech recognition device according to the present invention comprises: a stationary vowel standard pattern storage means for storing only a standard pattern of a stationary part of a vowel; The standardized vowel standard pattern and the input word standard pattern are compared with each other to identify a stationary vowel section in the input word standard pattern, and the word standard pattern correcting means for correcting the word standard pattern; Adaptive pattern matching means for adaptively performing linear and non-linear pattern matching on an input word pattern using the word standard pattern corrected by the method described above, and linearly matching consonant sections in the input word pattern. It is characterized in that matching is performed on a vowel section by nonlinear matching.

Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing a configuration of a word speech recognition apparatus to which a pattern matching method according to one embodiment of the present invention is applied.

The voice input unit 1 is a unit that digitizes a voice signal input through a microphone or the like. The start / end detection unit 2 is a unit that separates only a voice section from a silent section before and after a voice signal input by the voice input unit 1. The acoustic analysis unit 3 is a unit that analyzes the audio signal separated by the start / end detection unit 2 and extracts a feature parameter effective for recognition. The stationary vowel standard parameter memory unit 4 is a unit that stores only a feature parameter sequence corresponding to a stationary part of a vowel. Standard pattern correction unit 5
Is a unit for referring to the standard vowel standard pattern stored in the standard vowel standard parameter memory unit 4 to identify and correct the normal vowel part of the word standard pattern. The word standard pattern memory unit 6 is a unit that stores the corrected word standard pattern. The adaptive pattern matching unit 7 is a unit that adaptively performs linear and nonlinear matching between a feature parameter sequence of an input speech signal and a word standard pattern in accordance with information described in the word standard pattern. The recognition result processing unit 8 is a unit that performs a word recognition process based on the evaluation value obtained as a result of the adaptive pattern matching.

 The operation of this embodiment will be described below.

First, it is necessary to register a word standard pattern before performing a recognition operation. In the pattern registration operation, first, the speaker utters an isolated vowel, and the utterance is input to the voice input unit 1 through a microphone or the like. Voice input unit 1
In, the input signal is converted from an analog signal to a digital signal and sent to the next unit. Start / end detector 2
The input signal sent to the is grouped in frame units every several tens of samples, and a determination as to a silent section or a voice section is made in frame units based on information such as the average power and the number of zero crossings in the frame. Here, the frame sections determined as the voice sections are sequentially sent to the acoustic analysis unit 3. In the next acoustic analysis unit 3, the transmitted speech frame data is subjected to cepstrum analysis or the like in frame units, and the obtained feature parameters and power information are sent to the standard pattern correction unit 5. The standard pattern correction unit 5 identifies the received power information for each frame and sends only the characteristic parameters of the frame section in which the power is almost constant and flat to the stationary vowel standard pattern memory unit 4, and outputs the stationary vowel standard pattern. The memory unit 4 stores the feature parameter sequence as a standard vowel standard pattern of the vowel.

Next, in the registration operation of the word standard pattern, when the speaker utters a word to be registered, the same processing is performed from the voice input unit 1 to the sound analysis unit 3. However, the information sent from the acoustic analysis unit 3 to the standard pattern correction unit 5 is only the feature parameters in frame units, and the power information is not sent when the word information pattern is registered. In the standard pattern correction unit 5, the sent feature parameters are collated with the standard patterns stored in the standard vowel standard pattern memory unit 4 which have already been registered. The matching method is to perform end-point-free DP matching with each standard vowel standard pattern over the entire feature parameter sequence of the received word speech, and to determine the frame interval in which the matching evaluation value is less than or equal to a certain threshold value and minimum. As a stationary vowel section of the word standard pattern, a stationary vowel flag is set at the end of the feature parameter. The above processing is repeated for each vowel standard pattern.
The process ends when the evaluation of the matching no longer satisfies the condition below the threshold. However, if the result of this matching is that the normal vowel sections overlap, the section with the smaller evaluation value is adopted. The word standard pattern for which the identification of the steady vowel section has been completed is sent to the word standard pattern memory unit 6 as a corrected word standard pattern and stored.

The recognition operation can be performed only after the above two operations. In the recognition operation, when a speaker utters a word to be recognized, a voice signal input to the voice input unit 1 through a microphone or the like is processed up to the acoustic analysis unit 3 in the same manner as at the time of registration. The feature parameter sequence calculated by the acoustic analysis unit 3 is sent to the adaptive pattern matching unit 7 during the recognition operation. In the adaptive pattern matching unit 7, the input feature parameter sequence is sequentially collated with the word standard pattern stored in the word standard pattern memory unit 6. In the pattern matching method, linear matching is performed one-to-one sequentially from the beginning of the feature parameter sequence of the input speech and the beginning of the feature parameter sequence of the word standard pattern. Is performed by counting the cumulative sum of the distance values as the evaluation value. However, in the section where the steady vowel flag is set in the feature parameter of the standard pattern, DP matching is performed between the input pattern and the standard pattern. That is, in the regular vowel section of the standard pattern, DP
In this method, a cumulative distance value between patterns is obtained by matching and direct matching that does not expand or contract the time axis in other sections, and is used as an evaluation value. The evaluation value for a certain word standard pattern obtained by the above processing is
Confirmation result processing unit 8 along with the ID code of the standard pattern
Sent to The recognition result processing unit 8 executes a recognition algorithm such as selecting three candidates from the smallest evaluation value based on the received pattern matching result, and transmits the result to an external unit such as a host computer. With the above procedure, word level recognition is performed.

[Effects of the Invention] As described above, the present invention often performs DP matching in a stationary vowel section in a word standard pattern and direct linear matching in other consonants and transition sections, and thus often occurs in conventional DP matching. For erroneous recognition of words having similar vowel sequences, it is possible to reduce the erroneous recognition by increasing the ratio of the consonant part to the evaluation value. That is, there is an effect that the recognition rate can be improved in the pattern registration type word recognition method as compared with the conventional method.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a word voice recognition device to which a pattern matching method according to an embodiment of the present invention is applied, and FIG. 2 is a block diagram showing a configuration of a conventional word voice recognition device. 1 ... Speech input unit, 2 ... Start / end detection unit, 3 ... Sound analysis unit, 4 ... Standard vowel standard pattern memory unit, 5 ... Standard pattern correction unit, 6 ... Word standard pattern memory unit, 7 ...... Adaptive pattern matching section, 8 ... Recognition result processing section, 9 ... Host computer, 10 ... Pattern matching section.

Claims (1)

(57) [Claims] 1. A stationary vowel standard pattern storage means for storing only a standard pattern of a stationary part of a vowel, and a standard vowel standard pattern stored in the stationary vowel standard pattern storage means is collated with an input word standard pattern. Using a word standard pattern correcting means for identifying a stationary vowel section in the input word standard pattern and correcting the word standard pattern, and using the word standard pattern corrected by the word standard pattern correcting means, Adaptive pattern matching means for adaptively performing linear and nonlinear pattern matching, wherein linear matching is performed for consonant sections in the input word pattern, and nonlinear matching is performed for vowel sections. Pattern matching method for the word speech recognition device.