METHOD OF COLLATING VOICE PATTERN IN VOICE RECOGNIZING APPARATUS

VOICE PATTERN MATCHING METHOD OF A VOICE RECOGNITION UNIT -

FIELD OF TECHNOLOGY

This invention relates to a pattern matching method in a voice recognition unit.

BACKGROUND TECHNOLOGY

In voice recognition, in order to reduce recognition processing or to enhance recognition speed, it has been proposed first to preliminarily select those words high in similarity and then to carry out main recognition at high accuracy for the preliminarily selected words; however, conventionally, using the data of the beginning of a word, a main selection was carried out upon completion of all of the preliminary selection, so that, even if the end of a word had arrived during the preliminary selection, the main selection could not be entered until completion of the preliminary selection, whereby, in particular, there was a disadvantage of requiring long processing time for a relatively short word.

The present invention has been made so as to obviate the drawbacks of the prior art as described above, and, in particular, with an objective to carry out recognition processing in a voice recognition unit having two functions of preliminary and main selections at high speed.

DISCLOSURE OF THE INVENTION

In order to attain the above-described objective, in a voice recognition unit, in which preliminary selection is carried out using a certain range of data at the beginning of an unknown voice and a main selection is carried out upon completion of voice production, the present invention is characterized in that, in the case when a voice block signal changes to enter into a voiceless interval during a preliminary selection based on information at the beginning of a word, the preliminary selection is terminated to move into a main selection processing, and, then, its result is output if it has lead to termination of voice production; whereas, if it again returns to a voice interval, the preliminary selection is carried out by interrupting the main selection or discarding the result.

BRIEF DESCRIPTION OF THE DRAWINGS

Figs. 1 through 3 are time charts for illustrating embodiments of a voice pattern matching method according to the present invention.

MOST PREFERRED FORM TO PRACTICE THE PRESENT INVENTION

To describe the present invention more in detail, modes of practice of the present invention will be described with reference to the attached drawings hereinbelow.

Fig. 1 is a time chart for illustrating a process in the case where no voiceless interval is included in a voice interval, in which (a) indicates a single word pulse, (b). a voice interval pulse, and (c) a process procedure, wherein voice interval pulse (b) is a pulse obtained by detecting a voice power signal with a predetermined threshold value and the single word pulse (a) is a pulse generated upon detection of a rest interval (generally, 200 - 350 ms) upon completion of the voice interval pulse. Besides, in (c), A is a word beginning data take-in section, B is a preliminary selection section, B is a main selection section, and D is a result output section, wherein, the preliminary section is carried out on the basis of n frames of data at the beginning, and, as a result, top-ranked 1 number of words are extracted from subject words in a library to a certain number of words, followed by outputting a result by carrying out the main selection processing for the extracted result. However, in this case, in the case of a word including a voiceless interval, such as "INPUT", and a relatively short word, such as a numeral, the time for preliminary selection can be an obstacle for the main selection, which has obstructed to carry out the recognition processing time at high speed.

Fig. 2 is a time chart for illustrating the processing in the case of inclusion of a voiceless interval; in the first place, immediately following taking in of voice data A over a certain range from the beginning of a word, its preliminary selection B is entered, and, if a voice end edge a of a voice interval pulse is produced while the preliminary selection is being executed, the preliminary selection is interrupted promptly, and, upon storing all of an intermediate result of the preliminary selection and of the word currently being processed, the main selection C for the interval X is carried out. In this main selection, in the first place, using the previous intermediate result of the preliminary selection, the main selection is carried out _on the basis of the result of the preliminary selection up to the interrupted word, and, thereafter, the main selection is carried out for the entire word to attain a final result (one having the highest degree of similarity). If a start edge b of the next voice interval has reached during the processing of this main selection, the processing of said main selection C is interrupted, and the result of the main selection is all rendered invalid. This also holds true in the case of waiting for an end edge of a single word pulse upon completion of the main selection,. When the edge of b has been appreciated in this manner, the main selection is rendered invalid, thereby returning to the state of the next preliminary selection to continue the processing from halfway. Thereafter, if the voice/voiceless interval is repeated, the same processing is carried out until after the termination of the preliminary selection. On the other hand, when the single word pulse (a) has terminated, a result D of the main selection is output at that point in time. Then, in the case where the preliminary selection has terminated at some point in time Be, then the subsequent processing becomes only the main selection based thereon, so that the main selection is carried out each time when the. start edge of a voiceless interval is encountered, thereby waiting for the termination of the single word pulse and for the outcome of the output result. If processed in this manner, even if the length of a word is short, the most efficient processing can be carried out, thereby permitting an optimal high-speed processing by the preliminary selection.

Fig. 3 is a time chart showing, a processing example in the case where the preliminary selection has terminated, though a voice interval continues for a relatively long period of time, in the case shown in Fig. 2. In this case, upon completion of a first preliminary selection B₁, a second preliminary selection B2 is carried out on the basis of the first preliminary selection (e.g., the first preliminary selection B₁ being carried out by a frames of the beginning of a word and the second preliminary selection B₂ being carried out by m frames from the beginning of the word with m being larger than n), and similarly as described above, if it is carried out in combination with the main selection for the presence/absence of voice, the time period for processing the main selection from the end of the word may be further shortened, thereby allowing to handle an increased number of words and also to further shorten the time period for processing the main selection.

As is obvious from the above description, in accordance with the present invention, in a matching method of a recognition unit for carrying out the main selection processing from the end of a word, it is possible to carry out the optimal high-speed processing for a word of any length. Furthermore, by utilizing the present invention in pattern matching processing, there is obtained an optimal high-speed processing for a word of any length, thereby allowing to speed-up the processing time from the end of a word and to enhance the recognition rate and allowing to minimize the effect on the processing speed even if the number of words in a library is increased.

INDUSTRIAL APPLICABILITY

As described in detail above, the voice pattern identifying method of the present invention may be applied for voice recognition.