JPH0412480B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a word speech recognition method for recognizing words by comparing input speech with a word dictionary in which phonemes are expressed.

(Constitution of Conventional Example and Problems thereof) FIG. 1 is a functional block diagram of an apparatus for executing an example of a conventional word speech recognition method. A conventional example will be explained with reference to FIG. In Figure 1, 1
2 is a parameter extraction unit that creates a time series of parameters from input speech; 2 is a probability calculation unit that calculates the probability density of a phoneme by comparing standard patterns of phonemes; 3
is a word recognition unit that performs segmentation, likelihood calculation, word similarity calculation, etc. for each phoneme. Also,
4 is a phoneme standard created in advance through preliminary experiments etc. that expresses the distribution of various parameters for each phoneme in the form of an average value (μ _i ) for each phoneme and a covariance matrix (Σ _i ) between various parameters. A phoneme standard pattern section 5 stores patterns, and a word dictionary section 5 stores a word dictionary in which all words to be recognized are expressed in symbol strings in units of phonemes. In the word dictionary, for example, the words "Satsuporo" and "Asahikawa" are written as "SAQPORO" and "ASAHIKAWA", respectively.

Next, the operation of the above conventional example will be explained. The input phoneme is analyzed by the parameter extraction unit 1 every 10 ms frame, parameters are extracted, and a parameter time series is created. The probability density calculation unit 2 compares the parameters obtained for each frame with the phoneme standard pattern, and calculates the probability density of the phoneme generated from the parameter values. Next, the word recognition unit 3 performs phoneme segmentation for each dictionary item according to the dictionary phoneme series that makes up the dictionary item, and then segments the phoneme according to the type of phoneme and the phoneme corresponding to the phoneme according to the following formula. The likelihood l of the segmented interval is calculated, and the similarity is determined as the average of the likelihoods of each phoneme in the dictionary entry. Here, let the phoneme be X, and the frame numbers at the start and end of the segmented section corresponding to X are N _s ,
When N _e and the value of each parameter in the n-th frame are C _o , the likelihood l _x of the phoneme X is defined by the following formula.

represents the probability density of phoneme i in φ _i (C _o ),
It is defined as Eq.

φ _i (C _o )=1/(2π) ^j/2 |Σ _i | ^1/2 exp [−1/2(C _o −〓 _i ) ^T _-1 〓 _i (C _o −〓 _i )] …… C: j parameters in one frame (Bertl) 〓 _i : Average value (vector) of the parameters of a certain phoneme i Σ _i : Covariance matrix In the formula, the range of the denominator i in dividing the probability density is the phoneme It depends on what X is. For example, when X is the phoneme A, the range of i is five vowels, A, E, I, O, and U. The word similarity L _M obtained above was determined for each dictionary item according to the formula, and the dictionary item with the maximum L _M was selected as a recognized word.

L _M = _NP 〓 ⁱ⁼¹ li/NP …… L _M : Similarity of the Mth word in the dictionary l _i : Likelihood of phoneme i in the dictionary phoneme sequence NP: Number of dictionary phonemes that have just been segmented If the original phoneme in the word to be recognized in the interval is α, then the probability density of the phoneme α is φα. On the other hand, when we assume a different type of phoneme β and let the probability density be φβ,
If the values of φα and φβ are close, or if φα is smaller than φβ, as a result, the similarity of words that should not be recognized becomes higher than the similarity of words that should be recognized, It had the drawback of misrecognizing words.

(Object of the Invention) The present invention is intended to eliminate the drawbacks of the above-mentioned conventional example, and aims to improve the accuracy of likelihood calculation and thereby improve the word recognition rate.

(Structure of the Invention) The present invention provides a word dictionary in which words to be recognized are expressed in post-recording columns of phoneme units, and standard patterns of each phoneme and voiced and unvoiced sounds expressed in the distribution form of acoustic parameters of each phoneme. In order to perform word recognition of input speech, the input speech is checked against each dictionary entry in the word dictionary, and the input speech is segmented for each phoneme according to the dictionary phoneme sequence that constitutes each dictionary entry. In a word speech recognition method that uses the standard pattern to calculate the likelihood for the segmented speech section, not only the standard pattern of phonemes but also the presence of speech are detected within the segmented speech section. The feature of this method is that the likelihood is determined using both the dictionary entry and the standard pattern of non-speech, and this likelihood value is used to determine the degree of similarity between the dictionary entry and the input speech to recognize the word.

(Description of Embodiment) An embodiment of the present invention will be described below with reference to FIGS. 2 and 3. In FIG. 2, the parameter extraction section 1, probability density calculation section 2, phoneme standard pattern section 4, and word dictionary section 5 are the same as those in the conventional example shown in FIG.

A voiced sound/unvoiced sound standard pattern section 6 and a distance calculation section 7 between voiced sound and unvoiced sound are not present in the conventional example shown in FIG.

The patterns stored in the standard pattern section 6 for voiced and unvoiced sounds represent the distribution of the parameters of voiced and unvoiced sounds in the form of average values of voiced sounds and unvoiced sounds, and a covariance matrix between various parameters, and are created in advance. I'll keep it. The parameter time series obtained by the parameter extraction unit 1 is the same as in the conventional example.

The operation of this embodiment will be explained. First, the parameter extraction unit 1 obtains parameters for each frame from the input audio, and then uses the values of the parameters to
The probability density calculation unit 2 calculates the probability density obtained from each phoneme standard pattern in the same manner as in the conventional example. moreover,
The distance calculation unit 7 between voiced and unvoiced sounds compares the above parameter values obtained for each frame with the standard patterns of voiced and unvoiced sounds, and calculates the voiced or unvoiced sound generated from the parameter values. Distances L _v (n) and L _U (n) are calculated for each frame according to the formula.

L _V (n)=(C _o −〓 _V ) ^T Σ ^-1 _v (C _o −〓 _V ) L _U (n)=(C _o −〓 _U ) ^T Σ ^-1 _U (C _o −〓 _U ) However, Cn: Parameter (vector) in the nth frame 〓 _V , 〓 _U : Average value (vector) of the parameters of voiced and unvoiced sounds Σ _V , Σ _U : Covariance matrix of parameters of voiced and unvoiced sounds Next, each For each dictionary item, perform segmentation of phoneme X according to the dictionary phoneme sequence that makes up that dictionary item, and calculate the likelihood l _x of the segmented interval corresponding to that phoneme X and that phoneme However, when calculating l _x , consider a term that represents the likelihood of whether the assumed phoneme is a voiced or unvoiced sound, and use the likelihood l' _x instead of l _x according to the formula.

However, Ψ _V + Ψ _U = 1 Figure 3 shows the temporal changes in φ _x and Ψ _u when /ZiKU/ (axis) and /SiKU/ (lay) are uttered. Focusing on the voiced /Z/ and unvoiced /S/ of the first phoneme of /ZiKU/ and /SiKU/,
The value of φ _x is almost the same for /Z/ and /S/, and it is difficult to distinguish between /ZiKU/ and /SiKU/. However, looking at the value of Ψ _U , we find that Ψ _u of /S/
The value of Ψ _u of /Z/ is lower than the value of /Z/, and it is easy to distinguish between /ZiKU/ and /SiKU/ using l′ _x obtained from the formula.

This embodiment has the advantage that a highly accurate likelihood can be obtained by performing a likelihood calculation using the distance between a standard pattern of voiced and unvoiced phonemes within a segmented interval.

(Effects of the Invention) The present invention has the above-described configuration, and provides the following effects. A highly accurate likelihood can be obtained by performing a likelihood calculation using the distance between a standard pattern of voiced sounds and unvoiced sounds within a segmented interval.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram of the functions of an apparatus used to implement a conventional word speech recognition method, and FIG. 2 is a schematic block diagram of the functions of an apparatus used to implement a word speech recognition method according to an embodiment of the present invention. FIG. 3 is a diagram showing the temporal changes in the probability density φ _x and the voicelessness term Ψ _u when /ZiKU/ and /SiKU/ are uttered. 6... Voiced sound, unvoiced sound standard pattern section, 7... Distance calculation section between voiced sound and unvoiced sound.

Claims (1)

[Claims] 1 Equipped with a word dictionary that describes the words to be recognized as symbol strings in phoneme units, and standard patterns for each phoneme and voiced and unvoiced sounds expressed as a distribution of the acoustic parameters of each phoneme, it can recognize words from input speech. In performing this, the input speech is checked against each dictionary entry in the word dictionary, the input speech is segmented for each phoneme according to the dictionary phoneme series that constitutes each dictionary entry, and the standard pattern of the phoneme is used to segment the input speech. When calculating the likelihood for the segmented speech section, the likelihood is calculated using not only the standard pattern of phonemes but also the standard pattern of voiced and unvoiced sounds within the segmented speech section. A word speech recognition method characterized in that a word is recognized by determining the similarity between a dictionary item and input speech using the likelihood value.