JPH0634176B2 - Voice segmentation device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech segmentation apparatus in speech recognition technology, and particularly to a speech segmentation apparatus applied to an apparatus for recognizing speech using a fine recognition unit such as a phoneme. Regarding

(Prior Art) Since a method of recognizing phonemes as a unit does not need to prepare a standard pattern of words and is suitable for large vocabulary recognition, many recognition methods have been tried conventionally. Of these methods, the most common method is to use a phoneme standard pattern to perform matching between an input speech represented as a vector time series and a word dictionary represented as a phoneme symbol series. Is.

(Problems to be Solved by the Invention) The characteristics of a phoneme can be described not only as a standard pattern expressed as a vector but also as a local characteristic on various acoustic parameters. Such a local feature expression is not always expressed by the description in the standard pattern. The standard pattern does not necessarily include information about phoneme-specific or phoneme-specific phenomena such as abrupt changes in energy, the presence of energy peaks and valleys, and abrupt changes in spectra. For this reason, the performance of phoneme recognition by the phoneme standard pattern is insufficient. The present invention realizes a highly accurate matching using a phoneme standard pattern, that is, a highly accurate phoneme segmentation means by considering a phenomenon peculiar to a phoneme or a phoneme change.

It is an object of the present invention to provide such high-precision segmentation means to enable recognition in phoneme units, and to realize large vocabulary recognition that cannot be performed by word-based recognition.

(Means for Solving Problems) Means provided by the present invention for solving the above problems is a speech segmentation apparatus for dividing an input speech into a known phoneme sequence, which is expressed as a sequence of vectors. Input speech pattern A = {a (j)}, and a set of phoneme standard patterns represented as a vector B = {b (k, n), n = 1, ..., N (k ), K = 1,
, K}, a standard pattern storage unit, and an input pattern A
The frame distance d (k, j) for the phoneme category k is calculated from the distance between a (j) and each vector of the set B of standard patterns as a measure showing the similarity of the vector a (j) at each time with respect to the phoneme category k. A distance calculator that calculates, a phoneme boundary feature description table that describes the acoustic features that appear at the boundary points from phonemes to phonemes for each phoneme pair, and the frame distances in the section corresponding to each phoneme over the entire input section are accumulated. The amount (D +) obtained by adding the amount (P) representing the probability that the feature specified by the phoneme boundary feature description table exists at each phoneme boundary point as a penalty to the amount (D).
Means for dividing the interval so that P) is minimized.

(Operation) The vector sequence of the input speech is A = {a (j), j = 1,
,, j}. Also, set the phoneme standard pattern set to B.
= {B (k, n)}. Here, k represents a phoneme category number. k = 1, ..., K. n means that there are a plurality of standard patterns in each category, and n = 1,
..., N (k).

The frame distance for each phoneme category k in each frame j of the input is d (k, j) = min [dist (a (j), b (k, n)] (1) where dist (·, ·) represents the distance between the two vectors, and any value such as Euclidean distance or city distance can be used.

As a method of classifying an input into a known phoneme sequence using the phoneme standard pattern, the following method based on dynamic programming can be used. If the length of the phoneme string is M and the frame distance for the m-th phoneme is d (m, j), then the segment that minimizes the amount of accumulated frame distance over the entire input section is obtained from the recurrence formula. be able to.

Initial condition g (0,0) = 0, g (0, j) = ∞; j ≠ 0 (2) Recurrence formula N (m, j) = m'g (m, j) satisfying the above equation is a cumulative amount of distance, and N (m, j) is a buffer for a back pointer indicating an optimum course. By tracing the pointer from N (m, j) at the end, the optimum route, that is, the optimum section is obtained.

In the present invention, data called a phoneme boundary feature description table is newly introduced in order to execute a segmentation using a local feature at a phoneme boundary, that is, a change from a phoneme to a phoneme.

T = {t (k1, k2)} Here, t represents a feature observed at the phoneme boundary from the phoneme k1 to the phoneme k2, and is given for each combination of k1 and k2.

The following recurrence formula can be used instead of the recurrence formula of (2) so that a penalty is given when the feature specified by the above table does not exist.

Here, P [t, j] is a function having a value 0 when the feature t exists in the frame j, and a predetermined penalty value otherwise.

FIG. 1 shows the principle of phoneme segmentation according to the present invention. The figure shows the case of the word / saN /, but it is assumed that the phoneme boundary feature description table is described as follows for two phoneme boundaries sa and a-N.

t (/ s /, / a /) ＝ “maximum of low-frequency energy change” t (/ a /, / N /) ＝ “maximum of spectrum change” It is examined whether these features exist in each frame. , The penalty value sequence p (m, j) shown in the figure is obtained. Eventually, phoneme segmentation is performed by the above recurrence formula calculation so that the place where the penalty value becomes 0 becomes the phoneme boundary.

(Embodiment) FIG. 2 is a block diagram showing an embodiment of an apparatus realizing the present invention. An input pattern A and a phoneme standard pattern set B are stored in the input pattern storage unit 1 and the phoneme standard pattern storage unit 2, respectively.
Is stored. The phoneme string storage unit 3 stores a phoneme string description of the input voice, and the table storage unit 4 stores a phoneme boundary feature table. 5 is a distance calculation unit,
Equation (1) is calculated, the distance value {d (k, j)} is output and stored in the distance buffer 6. Reference numeral 7 denotes a local feature extraction unit, which extracts maximum and minimum frames of total band energy, band limited energy, spectrum change, etc. and stores them in the feature buffer 8. A recurrence formula calculation unit 9 executes the recurrence formula calculation of the formula (3) using the distances and features read from the distance buffer 6 and the feature buffer 8, and obtains the accumulated distance amount g, The pointer value N and the penalty value p for traceback of the optimum path are written in the buffers 10, 11 and 12, respectively. This operation is j = 1 to j
= J, and finally when j = J, the traceback unit 13 reads the contents of the buffer 11 and outputs the segmentation result.

(Effects of the Invention) As described above, according to the present invention, in the segmentation by the dynamic programming method using the phoneme standard pattern, the local feature information of the phoneme that is not always sufficiently described in the standard pattern. , It is possible to realize highly accurate phoneme segmentation. The present invention can be applied to the realization of large vocabulary recognition using a phoneme as a recognition unit.

[Brief description of drawings]

FIG. 1 is a diagram for explaining the principle of the present invention, and FIG. 2 is a block diagram showing an embodiment according to the present invention. In the figure, 1, 2, 3, and 4 are storage units, 5 is a distance calculation unit, 6, 8, 10, 11, and 12 are buffers, 7 is a local feature extraction unit, 9 is a recurrence formula calculation unit, and 13 is a trace. It is the back part.

Claims (1)

[Claims] 1. A speech segmentation apparatus for classifying an input speech into a known phoneme sequence, and an input pattern storage section for storing an input speech pattern A = {a (j)} represented as a series of vectors, and a representation as a vector. Set of phoneme standard patterns to be generated B = {b (k, n), n = 1, ..., N (k),
A standard pattern storage unit that stores k = 1, ..., K} and a phoneme category k of a vector a (j) at each time of the input pattern A.
A distance calculation unit that calculates a frame distance d (k, j) for the phoneme category k from the distance between a (j) and each vector of the set B of standard patterns as a measure indicating the similarity to
A phoneme boundary feature description table in which acoustic features appearing at boundary points from phonemes to phonemes are described for each phoneme pair, and an amount (D) in which the frame distance in a section corresponding to each phoneme is accumulated over the entire input section, And a means for dividing the section so that the amount (D + P) obtained by adding the amount (P) representing the probability that the feature specified by the phoneme boundary feature description table exists as a penalty at each phoneme boundary point is minimized. An audio segmentation device characterized by the above.