JP3110025B2 - Utterance deformation detection device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an utterance deformation detection device that automatically detects utterance deformation occurring in uttered speech with high performance.

(Prior Art) Conventionally, in a method of recognizing an unknown voice, a plurality of standard models created from voice data uttered in advance, that is,
In general, a similarity between a standard pattern and a voice pattern obtained from an unknown voice is determined, and a category of a standard model that gives the maximum similarity is used as a recognition result. Therefore,
Of course, the voice data for creating the standard pattern must have known utterance contents, and this becomes more strict when a unit smaller than a word such as a phoneme is used as a recognition unit.

On the other hand, even when the same word is uttered, depending on the type of the word, utterance deformation such as silence or long vowels may occur. For example, the two monophonemes / u / of "applause" / hakusyu / may be vocalized and voiced, and "/ eiga / of movie" / eiga /
i / may be uttered as a long vowel as / ee /. As a result, when compared on a phoneme basis, the word names indicate that the phonemes are the same, but have different voice patterns. Hereinafter, the “phoneme” means not only the minimum basic unit of speech in the phonological sense, but also a wider range of speech units including syllables and concatenation of a plurality of phonemes. As a method for detecting utterance deformation, for example, Takeda,
Kagiri et al., In a paper entitled "Phonological Labeling for Speech Database Construction" (hereinafter referred to as Reference 1) published on pages 69-70 of the Transactions of the Acoustical Society of Japan Spring Meeting 1987, pp. 69-70. Such an inspection method may be used. Here, phonetic labeling of the voice is performed based on a spectrogram, a waveform, and the like of the voice data, and at that time, a vocal deformation is detected.

(Problems to be Solved by the Invention) When the detection of vocal deformation as described in the above-mentioned conventional technique is performed by inspection as described in Document 1, if the number of voice data becomes enormous, a great deal of work is required. turn into. There is also a problem that the detection result may be different depending on the person who performed the detection work.

SUMMARY OF THE INVENTION An object of the present invention is to provide an apparatus for automatically detecting, with high performance, an utterance deformation occurring in an uttered sound, excluding the above-described drawbacks.

(Means for Solving the Problems) In order to solve the above-mentioned problems, an utterance deformation detection device according to the present invention analyzes a voice signal and outputs a feature vector time series, and outputs a utterance deformation of the voice signal. A vocal transformation information detecting unit that extracts a plurality of phonation transformation patterns assumed from among phoneme names of phonemes that may possibly occur and utterance transformation patterns stored in advance and outputs them as utterance transformation information, A standard model storage unit that stores a model in advance, and a phoneme name of a phoneme that may be uttered and deformed based on the feature vector time series, the utterance deformation information, and the standard model stored in the standard model storage unit. A deformed phoneme extraction unit that extracts position information of a phoneme that may be uttered and outputs it as deformed phoneme information, and a deformed phoneme information storage unit that stores the deformed phoneme information Based on the feature vector time series, the deformed phoneme information stored in the deformed phoneme information storage unit, and the standard model stored in the standard model storage unit, a plurality of utterance deformations extracted in the utterance deformation information detection unit. An utterance deformation detecting unit that determines which utterance deformation occurs in the pattern.

(Operation) Hereinafter, the operation of the utterance deformation detection device according to the present invention will be described.

The present invention cuts out a speech section for a phoneme that may be uttered and deformed (hereinafter, referred to as a deformable phoneme) from a uttered input speech, and inputs a speech model from a standard model for the phoneme and a speech pattern of the section. The presence or absence of the utterance deformation of the utterance is automatically detected.

In order to detect the utterance deformation of the input voice, first, a generated deformation pattern that may occur with respect to the utterance content must be obtained. Many utterance deformations, particularly utterance deformations due to allophones, can be ruled for the ease of occurrence of the deformation depending on contexts of the phonemes before and after. "Voiceless consonants, vowels / i /, / u / sandwiched between endings are easily voiced."
ei /, / ou / are each easily converted into long vowels into / ee /, / oo /, respectively. Then, a pattern created by these rules or another pattern that is empirically known to cause utterance deformation is defined as an utterance deformation pattern.

In order to determine the position of the deformable element in the input voice, for example, a standard model (hereinafter, referred to as a phoneme unit) based on a plurality of voice deformation patterns corresponding to the voice content of the input voice is used.
A model for each pattern (hereinafter referred to as a deformation model) is created by linking the phoneme models (called a "transformation model"). Two). As phoneme models, for example, SElevinson, LRRabiner, and MMS
ondhi et al., The Bell System Technical Journal, Vol.
62, No. 4, April 1983, pp. 1035-1074, “An Introduction to the Application of the Theory”
of Probabilistic Functions of Markov Process to A
utomatic Speech Recognition "(hereinafter referred to as Reference 2)
Hidden Markov model (hereinafter HM)
M). The HMM is a type of state transition network, and a state transition probability and a vector appearance probability are defined for each state. Then, the parameters of the HMM are obtained by using a learning voice, and a forward-backward as described in Reference 2.
Estimated by algorithm.

After creating the deformed model, the position of the deformable phoneme is calculated for each model using the feature vector time series obtained from the input voice. (For "movie", the positions of / ei / and / ee / become). Here, as a method of obtaining the feature vector time series, for example, the method described in pages 154 to 160 of “Digital Speech Processing” (hereinafter referred to as Reference 3) by Furui, published in 1985 by Tokai University Press. A cepstrum method, an LPC analysis method, or the like can be used.

In addition, the position of a deformable phoneme with respect to a certain deformation model can be determined, for example, by using Viterb (Viterb
i) An optimal state transition path in the model is determined using an algorithm, and a section in the input speech corresponding to a path of a phoneme model (hereinafter referred to as a deformable phoneme model) for a deformable phoneme on the path. Is required.

For example, the presence or absence of vocal deformation is determined by a forward-backward algorithm or a Viterbi algorithm for each deformable phoneme model (in the case of "movie", a phoneme model for / ei / and / ee /). First, the appearance probability of the voice pattern of the voice section obtained earlier is obtained, and the model having the highest probability can be determined as the phoneme of the section.

(Embodiment) Next, an embodiment of an utterance deformation detection device according to the present invention will be described with reference to the drawings.

 FIG. 1 is a block diagram showing one embodiment of the present invention.

A phoneme model M using an HMM as described in Literature 2 is stored in the standard model storage unit 3.
These can be created from a large amount of audio data by the forward / backward algorithm described in Reference 2.

The input audio signal S is input to the feature analysis unit 1 and the utterance deformation information detection unit 2.

The feature analysis unit 1 converts the audio signal S into a feature vector time series V by using a mel-cepstral method as described in Document 3.

The utterance deformation information detection unit 2 obtains an utterance deformation pattern that may occur with respect to the utterance content of the voice signal S according to the rule, and outputs the utterance deformation pattern along with the deformable phoneme name in the utterance content as utterance deformation information P. .

In addition to this method, the utterance deformation pattern can also be obtained by preparing in advance a memory covering all the utterance deformation patterns of all the inputted voices and extracting a necessary pattern from the memory.

The deformed phoneme extraction unit 4 receives the feature vector time series V, the utterance deformation information P, and the phoneme model M stored in the standard model storage unit 3, and connects the phoneme model M for each utterance deformation pattern to generate a deformation model. The optimal state transition path for the feature vector time series V is determined using the Viterbi algorithm described in Document 2 and the feature vector time series corresponding to the path occupied by the deformable phoneme model on that path. The start and end of the section in V are obtained as the position information of the deformable phoneme, and output as the deformed phoneme information I together with the deformable phoneme name.

The deformed phoneme information I is stored in the deformed phoneme information storage unit 5.

The utterance deformation detecting unit 6 receives the deformed phoneme information I ′, the feature vector time series V of the input speech signal, and the phoneme model M stored in the deformed phoneme information storage unit 5, and receives the deformable phoneme name in the deformed phoneme information I ′. The appearance probability of the subsequence of the feature vector time series V extracted based on the position information of the deformable phoneme in the deformed phoneme information I ′ for each model for each phoneme model corresponding to the forward and backward algorithm , The model with the highest probability is determined to be a phoneme in that section, and the phoneme name R is output as a detection result.

(Effects of the Invention) As described above, the present invention cuts out a speech section for a phoneme that may be uttered and deformed with respect to an input speech, and transforms the utterance of the input speech from a standard model for the phoneme and a speech pattern of the section. Since the presence / absence of the utterance is automatically detected, it is possible to realize a high-performance utterance deformation detection device which reduces the work of the detector.

[Brief description of the drawings]

 FIG. 1 is a block diagram showing one embodiment according to the present invention. 1 ... Analysis unit, 2 ... Transformation information detection unit, 3 ... Standard model storage unit, 4 ... Transformation phoneme extraction unit, 5 ... Transformation phoneme information storage unit, 6 ... Transformation detection unit.

Continuation of front page (56) References JP-A-63-5395 (JP, A) JP-A-63-205699 (JP, A) JP-A-1-126694 (JP, A) The Bell System Technical Journal Vol. . 62, No. 4, April 1983, p. 1035-1074

Claims (1)

(57) [Claims] 1. A feature analysis unit for analyzing a speech signal and outputting a feature vector time series, a phoneme name of a phoneme which may possibly cause speech transformation of the speech signal, and a phoneme name presumed from a speech transformation pattern stored in advance. An utterance deformation information detecting unit that extracts a plurality of utterance deformation patterns to be output and outputs the utterance deformation pattern as utterance deformation information
A standard model storage unit that stores a standard model in units of phonemes in advance; and the possibility that the utterance deformation is performed based on the feature vector time series, the utterance deformation information, and the standard model stored in the standard model storage unit. A deformed phoneme extracting unit that extracts a phoneme name of a certain phoneme and positional information of a phoneme that may be uttered and outputs as deformed phoneme information, a deformed phoneme information storage unit that stores the deformed phoneme information, and the feature vector Based on the time series, the deformed phoneme information stored in the deformed phoneme information storage unit, and the standard model stored in the standard model storage unit, any one of the plurality of utterance deformation patterns extracted by the utterance deformation information detection unit And an utterance deformation detecting unit for determining whether or not the utterance deformation has occurred.