JPH0254560B2 - - Google Patents

Description

[Detailed Description of the Invention] (1) Technical Field of the Invention The present invention provides high-speed and high recognition accuracy in real-time speech recognition processing using a multi-word speech recognition method, which includes a pre-verification processing unit that selects candidate words. The present invention relates to a speech recognition device having the following.

(2) Prior art and problems Conventionally, pre-matching processing methods in multi-word speech recognition devices include those that extract global features of the spectrum, those that perform time normalized matching of normalized power, or methods that extract details of the beginning and end of words. There are methods based on pattern matching. Word pronunciation length, which is a basic difference between words, or power information of vowels and consonants is not actively applied to these words. Therefore, the parameters in the pre-verification are essentially the same as those in the main verification, and there is a drawback that quite detailed information must be used in order to improve the classification function, and the verification takes time.

(3) Purpose of the Invention The purpose of the present invention is to identify word utterance lengths or vowels that indicate basic differences between words in a multi-word speech recognition system.
It is an object of the present invention to provide a speech recognition device having high speed and high recognition accuracy by applying power amount information of consonants.

(4) Structure of the Invention In order to achieve the above object, the speech recognition device of the present invention is a speech recognition device having a pre-verification processing unit that performs real-time recognition processing of speech using a multi-word speech recognition method and selects candidate words. In the pre-verification processing unit, the word utterance length, the total power amount of the word utterance obtained from the normalized power, and the time length with relatively high power and the time length with relatively low power indicating vowel nature and consonant nature are provided. The present invention is characterized by providing means for selecting candidate words using three feature parameters having a ratio of .

(5) Embodiments of the invention FIG. 1 and FIGS. 2a and 2b are explanatory diagrams of the main parts of the present invention.

As a method for selecting candidate words in the multi-word speech recognition device of the present invention, the parameters of the pre-matching process are set separately from the identification parameters of the main matching process.
Three quantities: the word utterance length, the total power of the word utterance calculated from the normalized power, the average value of the normalized power plus the standard deviation value as a quantity indicating vowelness and consonance, and from this Candidate words are selected by using three feature parameters: subtract and ratio.

FIG. 1 is an explanatory diagram regarding time normalization used in this case.

As shown in waveforms 1 ₁ to 1m in the figure, the utterance time length of word sounds differs not only for different words but also for each utterance of the same word.

Therefore, as shown in waveform 2 in the figure, it is normalized to a reference time length, and this time-normalized verification waveform is used for verification with the dictionary. At this time, dictionaries to be checked that have extremely different lengths, that is, dictionaries whose length is twice or more or half or less, are excluded. In the conventional system, this time length normalization, which is an inherent quantity of word utterance, was not actively applied, but this is set in the present invention.

FIGS. 2a and 2b are explanatory diagrams of the total power amount of word utterance used in the pre-verification processing section of the present invention.

Figures a and b show power changes in the time direction of word utterances when the horizontal axis is the time length and the vertical axis is _the power.
3 ₂ is shown as an example, and corresponds to the same time-normalized word shown in FIG.

Normally, even if the word is the same, not only the time component but also the magnitude of the power will differ depending on how it is uttered.
These word waveforms 3 ₁ , 3 ₂ are linearly normalized between the maximum power and the minimum power to form word waveforms 4 ₁ , 4 _{2 .}
is obtained. Note that although the time length of the uttered word varies depending on the utterance, there is a rough value that is unique to the word. Therefore, as shown in the figure, the total power of the word can be obtained from the shaded portions of the word waveforms 4 ₁ and 4 ₂ by normalizing the power along with the time length as a relatively simple amount of specific information of the word.

With the above method, the first power of the three feature parameters is the normalized vocal utterance length, and the second parameter is the sum of the normalized utterance parameters over the entire time length, that is, the total power of the word. It's the amount. The total amount of power shown in FIG. 2, which is a combination of these two parameters, is set as a stable parameter as the amount of unique information of word utterance.

The next third power is a value of vowel volume/consonant volume during word pronunciation as an index indicating vowel-likeness and consonant-likeness in a word. The length of time during which the normalized power exceeds (its average value) + (standard deviation) is used as the vowel volume, and the length of time during which the normalized power exceeds (its average value) - (standard deviation) is used as the consonant volume. long is used. In other words, the index is (vowel/consonant ratio) = (average value) + (standard deviation) / (average value) - (standard deviation) [normalized power] x 100. This becomes an effective parameter for word classification.

FIG. 3 is an explanatory diagram of the configuration of an embodiment of the present invention.
Prior to recognition, a pre-verification dictionary 16 and a main verification dictionary 17, which will be described later, are prepared. In the pre-verification dictionary 16, words are classified in the order of the magnitude of each value regarding three feature parameters, and in the pre-verification, one parameter value of the input voice is determined, and dictionary items that fall within ±30% of that value are selected. do. This is done for the three parameters, and the logical product of the three is taken to select a candidate word as the pre-verification result.

The main matching dictionary 17 stores normal feature parameters such as spectral patterns used in the main matching.

In the figure, during recognition, voice is input from a microphone 10, and the electrical signal is sent to an analysis circuit 12 through an amplifier 11, where it analyzes various parameters necessary for voice recognition.

First, for pre-matching processing to select candidate words, the power normalization circuit 14 normalizes the speech power using the method shown in FIGS. / consonant ratio. These are the pre-verification circuit 1
5, the word is checked against the pre-check dictionary 16, and candidate words are selected as described above. This is sent to the main reference dictionary 17, and a corresponding parameter such as a spectrum pattern is selected. On the other hand, in the main matching process, in order to extract complementary quantities to the parameters used in the pre-matching process, the output of the analysis circuit 12 is sent to the feature extraction circuit 13, and, for example, feature parameters of the spectral pattern are extracted, and the main matching circuit 18
In this step, the distance between the parameter and the parameter from the reference dictionary 17 is calculated, and the result is sent to the determination circuit 19, which makes a determination and outputs the recognition result.

(6) Effects of the Invention As explained above, according to the present invention, in the pre-verification processing section, the above-mentioned three characteristic parameters based on the normalized power of the word as the amount of information specific to the word utterance are used to Since the matching processing section performs matching using complementary parameters and also performs two-stage matching processing, candidate words can be selected stably and quickly, and word speech recognition can be performed with a high recognition rate.

[Brief explanation of the drawing]

1 and 2 a and b are explanatory diagrams of the main parts of the present invention, and FIG. 3 is an explanatory diagram of the configuration of an embodiment of the present invention, 4 ₁ and 4 ₂ are normalized total power amounts, 10 is a microphone, 11 is an amplifier, 12 is an analysis circuit, 13 is a feature extraction circuit, 14 is a power normalization circuit, 15 is a pre-verification circuit, 16 is a pre-verification dictionary, 17
Reference numeral 18 indicates a book checking dictionary, 18 a book checking circuit, and 19 a determination circuit.

Claims (1)

[Claims] 1. In a speech recognition device that has a pre-verification processing unit that performs real-time recognition processing of speech using a multi-word speech recognition method and selects candidate words, the pre-match processing unit is configured to perform real-time speech recognition processing using a multi-word speech recognition method to select candidate words. A means is provided for selecting candidate words using three characteristic parameters: the total power of word utterance, and the ratio of the time length with relatively high power and the time length with relatively low power indicating voweliness and consonance. A voice recognition device featuring: