JP3061292B2 - Accent phrase boundary detection device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used in a continuous speech recognition apparatus for recognizing a continuously uttered speech. The present invention relates to an accent phrase boundary detection device that detects a.

[0002]

2. Description of the Related Art Conventionally, in detecting an accent phrase boundary of a continuous speech, analysis of a feature parameter (for example, pitch frequency, power, etc.) pattern related to the prosody included in the speech is performed based on the analysis of the prosody feature parameter time series pattern. A technique for detecting this valley using the fact that many phrase boundaries exist has been proposed. (For example, Suzuki et al., "Phrase Boundary Detection Using Prosody Information for Japanese Continuous Speech Recognition," IEICE Transactions, Vol. J72-
D-2, no. 10 (1989-10); Komatsu et al., "Syntax Estimation Using Prosody Information and Conversational Speech Understanding Method Using Word Spot", Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J
71-D, No. 7 (1988-7)) FIG. 5 shows a pitch frequency pattern when one sentence is read out.
The vertical axis is pitch frequency, and the horizontal axis is time. Dashed lines are the boundaries of accent phrases. In some places, the portion where the envelope is interrupted is a portion in which a pitch frequency cannot be obtained, such as a pause interval between unvoiced consonants and utterances. Since the pitch frequency pattern is quite vague, there are many variations (variations), and there are sections where the pitch frequency cannot be extracted, it is very difficult to detect the valley of the envelope. However, the conventional methods scan the prosodic feature parameter patterns locally, or detect accent phrase boundaries based on heuristic or empirical rules, and have insufficient detection performance and lack versatility. .

[0003]

According to the present invention, a hidden Markov model (for example, Seiichi Nakagawa, "Speech Recognition by Stochastic Model" (1988)) is used for each accent type of a phrase and for each duration of the entire phrase. )), A statistical standard pattern for statistically expressing the characteristic parameter time-series pattern of the accent phrase is stored in the statistical accent phrase model memory in advance, and the characteristic of the input speech is detected by the phrase boundary detection unit. The parameter time series and the statistical standard pattern are collated while controlling the length of stay of the internal state of the hidden Markov model, and the parameter time series pattern of the accent phrase in the continuous speech is recognized one by one, and the optimal A model (standard pattern) sequence is obtained, and the boundary of each model is set as an accent phrase boundary.

In other words, according to the present invention, the time-series pattern of the characteristic parameters relating to the prosody in each interval of the accent phrase is changed depending on the accent type (for example, "Japanese Pronunciation Accent Dictionary" edited by Japan Broadcasting Corporation (1966)). We focus on the fact that it can be classified into typical patterns.
1A and 1B show typical examples of each accent type and its pitch frequency pattern. Accent type is defined by the number of mora from the beginning of the word to the syllable with the accent nucleus (hatched in the figure), and the pitch frequency pattern drops sharply immediately after the syllable with the accent nucleus. In FIG. 1A, the key at the left end of the horizontal line indicates that the pitch frequency increases, and the key at the right end indicates that the pitch frequency decreases. A time series pattern of prosodic feature parameters in accent phrases is classified for each accent type, and a statistical accent phrase model is created for each accent type by a hidden Markov model using a time series pattern of feature parameters of a large number of accent phrases. . This model statistically represents a time-series pattern of feature parameters of an accent phrase.
As described above, detecting the accent phrase boundary by comparing the special sequence pattern of the prosodic feature parameter of the input continuous speech with the statistical accent phrase model requires statistical standard pattern processing in order to handle ambiguous prosodic parameters. This means that the accent phrase boundary is detected by globally considering the parameter time series of the input voice, which is represented by a model, and the detection performance is improved as compared with the related art.

[0005]

FIG. 2 shows an embodiment of the present invention. The audio input from the input terminal 1 is converted into a digital signal in the A / D converter 2. After the digital signal is pitch-extracted by the pitch frequency extracting unit 3, it is further converted into characteristic parameters every frame (for example, every 10 milliseconds). This characteristic parameter is, for example, one of the pitch frequency patterns of several adjacent frames (for example, seven frames).
It is a secondary regression coefficient or a secondary regression coefficient.

After converting a large amount of a learning sentence database into the above-mentioned feature parameters, a feature parameter time series is cut out for each accent phrase and classified for the same accent type. Further, for each duration of the accent phrase (for example, 400 ms or less, 400 ms to 600 ms, 60 ms or less).
0 ms or more), and a statistical accent phrase model is created using a hidden Markov model represented by several states (for example, three states). Phrases having extremely short durations and phrases having extremely long durations have different characteristics, so classification by time length is effective. These models are stored in the statistical accent phrase model memory 4.

The phrase boundary detecting unit 5 reads out a statistical accent phrase model from the statistical accent phrase model memory 4 and compares it with the feature parameter time series of the input speech.
Find out which model series matches best. At the time of matching, each state of the Hidden Markov Model used for the statistical accent phrase model is subjected to duration control. That is,
The average time length and variance remaining in each state are previously obtained from the learning data, and the normal distribution is assumed for the stay time of the state, and the likelihood is obtained from the actually matched duration time,
Add to the likelihood of matching. For example, when a statistical accent phrase model is created using a three-state hidden Markov model, each state includes a rising part of the pitch pattern of the phrase, a flat part,
It is thought to be created to represent the descending part. FIG.
Shows the average duration for each accent type in each state when the statistical accent phrase model is created by a three-state hidden Markov model. The duration of the second state, that is, the state of mapping to the flat part of the pitch pattern, increases from type 2 to type 4. As can be seen from the comparison with FIG. 1B, this is consistent with the tendency of the actual pitch pattern. By adding the duration control, the features of each accent-type pitch pattern can be reflected in the model.

In the thus obtained optimal model sequence for the input speech, the boundaries of the respective models correspond to the accent phrase boundaries, which are examined and output from the phrase boundary detection result output unit 6.

[0009]

As described above, in the present invention, since the prosodic feature parameters are modeled by a statistical method, many variations of the prosodic feature parameter time-series patterns are stochastically represented as standard patterns. be able to. Therefore, it is possible to cope with a pattern that changes subtly in the continuous voice, and the detection performance is improved.

According to the configuration shown in FIG. 2, a male announcer utters 500 sentences (utterance speed 10 mora /
The following shows the experimental results of accent phrase boundary estimation using the First, based on 400 sentences (including 2,786 accent phrases), statistical accent phrase models of type 0 to type 4 were created. At this time, all type 4 or more accent phrases were used to create the type 4 model. Also, according to the duration of the entire phrase, 400 ms or less,
It was classified into three categories of 400 ms to 600 ms and 600 ms or more. Therefore, a total of 15 statistical accent phrase models are created. As characteristic parameters, a first-order regression coefficient and a second-order regression coefficient of a pitch frequency pattern were used, and a hidden Markov model having three states was used. Next, a phrase boundary detection evaluation experiment was performed on 100 sentences (including 463 accent phrase boundaries) different from the above. Under the condition that the number of accent phrases in one sentence and the accent type of the accent phrase are known, when the same number of boundary candidates as the number of correct phrase boundaries are detected, the detection boundary error is detected within ± 100 ms (about 1 mora length). The border ratio is 76.
4%. Further, when the duration control is added to the statistical accent phrase model, it is improved to 82.9%. FIG.
An example of the accent phrase boundary detection result according to the present invention is shown below. The vertical axis is pitch frequency, and the horizontal axis is time. The dashed line indicated by the upper arrow indicates a phrase boundary based on the phoneme label attached by the inspection, and the dashed line indicated by the lower arrow indicates the estimated phrase boundary. Accurate phrases are estimated with good accuracy except for the boundaries of phrases in which the duration of the accent phrase is short and sufficient prosodic information cannot be obtained.

If a power pattern is used as a feature parameter, the performance is further improved.

[Brief description of the drawings]

FIG. 1A shows an example of accent types 0 to 5;
B is a diagram showing these pitch frequency patterns.

FIG. 2 is a block diagram showing an embodiment of the present invention.

FIG. 3 is a diagram showing, for each accent type, the average duration of each state when a statistical accent phrase model is created using a three-state hidden Markov model.

FIG. 4 is a diagram showing an example of a phrase boundary detection result of continuous speech according to the present invention.

FIG. 5 is an example of a pitch frequency pattern of continuous speech, and the utterance content is “speech / freedom / one step / one yield / one hundred steps / one thousand steps / attacked”.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-64-49098 (JP, A) JP-A-64-20599 (JP, A) JP-A-62-34200 (JP, A) JP-A 62-34200 102295 (JP, A) JP-A-4-66999 (JP, A) JP-A-62-136699 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 15/04 G10L 15 / 14 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. An accent phrase boundary detecting device for converting a continuously uttered input speech into a feature parameter time series and detecting a boundary of an accent phrase contained in the input speech from the feature parameter time series, A statistical accent phrase that stores a statistical standard pattern that statistically expresses the feature parameter time-series pattern of the accent phrase created using a hidden Markov model for each accent type and for each duration of the entire phrase A model memory, and a phrase boundary detecting unit that detects the accent phrase boundary by matching the feature parameter time series of the input speech and the statistical standard pattern while controlling the length of stay of the internal state of the hidden Markov model. An accent phrase boundary detecting device provided. 2. The voice characteristic parameter is a linear regression coefficient or a secondary regression coefficient of a change pattern over several frames of a pitch frequency of the voice, and any one of them is combined with a waveform power of each frame. The accent phrase boundary detecting device according to claim 1, wherein