JP3033132B2 - Language processor - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a language processing apparatus using a dictionary search method by DP matching.

[Prior Art] Conventionally, as one method of performing a dictionary search from phoneme recognition results including ambiguity such as voice input, there is a dictionary search method by DP matching, and this dictionary search method by DP matching is used. The language processing device includes a dictionary in which phonemes are written, and word extracting means. The word extracting means extracts words by DP matching for each word in the dictionary and outputs word candidates.

[Problems to be Solved by the Invention] However, in the conventional dictionary search method using DP matching, a large vocabulary dictionary of about 40,000 words for a word processor is used in order to match all words in the dictionary. However, there is a problem that the processing time becomes long.

The present invention has been made to solve the above-described problems, and has as its object to provide a language processing apparatus with a short processing time.

Means for Solving the Problems In order to achieve this object, a language processing apparatus according to the present invention uses an input means for inputting linguistic information by voice, analyzes the input linguistic information, and outputs a phoneme based on the analysis result. Phoneme recognition means for performing recognition; search means for searching for a word candidate from a phonemic notation dictionary based on the phoneme sequence recognized by the phoneme recognition means; and text based on the word candidate searched for by the search means. The word candidate searched by the search means is a word composed of less than a predetermined number of phonemes or a word composed of more than a predetermined number of phonemes. And a word extraction by DP matching is performed on the word candidates determined to be composed of less than a predetermined number of phonemes by the phoneme number determination means. The phonetic word extracting means and the phoneme number determining means determine that the partial word composed of less than the predetermined number of phonemes present in the word candidate determined to be composed of the predetermined number or more of the phonemes is Word extraction determining means for determining whether or not a word has been extracted by the first phoneme word extracting means; and if the partial word has been determined to have been extracted by the word extraction determining means, the partial word And a second phoneme word extraction unit that performs word extraction by DP matching on a word candidate including the word candidate. If the word extraction determination unit does not determine that the partial word has been extracted, the partial word is extracted. The second phoneme word extracting means does not perform DP matching on the included word candidates.

[Operation] In the language processing apparatus of the present invention having the above configuration,
Linguistic information by voice is input by the input unit, the input linguistic information is analyzed by the phoneme recognition unit, phoneme recognition is performed based on the analysis result, and retrieval is performed based on the phoneme sequence recognized by the phoneme recognition unit. A word candidate is searched from the dictionary in which the phoneme is written by the means. The phoneme number determination means determines whether the word candidate searched by the search means is a word composed of less than a predetermined number of phonemes or a word composed of a predetermined number or more of phonemes. The first phoneme word extracting means selects a word candidate determined to be composed of less than a predetermined number of phonemes.
Extract words by DP matching. The word extraction determining means is configured to determine whether the partial word composed of less than the predetermined number of phonemes present in the word candidates determined to be composed of the predetermined number or more of the phonemes by the phoneme number determining means is the first word. It is determined whether or not the word has been extracted by the phoneme word extracting means. If the word extraction determining means determines that the partial word has been word extracted, the second phoneme word extracting means DP for word candidates containing words
If word extraction is performed by matching, and the word extraction determining means does not determine that the partial word has been word extracted, DP matching is performed on the word candidate including the partial word by the second phoneme word extracting means. Not performed. Then, the output processing means outputs a sentence from the searched word candidates.

[Embodiment] An embodiment of the present invention will be described below with reference to the drawings.

First, the configuration of a language processing apparatus to which the present invention is applied will be described with reference to FIG.

A RAM 27, a ROM 28, and an I / O port 25 are connected to the CPU 26, respectively, and a display 29 is provided on the I / O port 25. Further, the I / O port 25 is provided with a microphone 21, an audio amplifier 22, a low-pass filter 23, and an A / D converter 24 continuously.

Next, the operation of the language processing apparatus of this embodiment will be described with reference to FIG.

In step S1 (hereinafter simply referred to as S1 and the same applies to other steps), the input continuous sound is converted into an electric signal by the microphone 21, and the audio amplifier 22
Amplified by This amplified signal is a low-pass
Only the components of 5.5 kHz or less are input to the filter 23 and pass therethrough, thereby suppressing aliasing at the time of sampling. This output signal is
The A / D converter 24 samples the data at 12 kHz and 16 bits. The sampled data is captured by the CPU 26 via the I / O port 25 and stored in the RAM 27 (voice capturing unit).

In S2, the CPU 26 reads the data from the RAM 27, and performs pre-emphasis by a digital filter having a 1-Z-1 transfer function. For this output data, 4 msec is regarded as one frame, and 21.3 msec per frame.
A long Hamming window is applied, and the 16th-order LPC cepstrum and power are calculated by LPC analysis (analysis unit).

In S3, the LPC calculated by the analysis unit S2 by the CPU 26
With the cepstrum and power as inputs, phoneme extraction is performed. Next, phoneme recognition is performed on the cut-out phoneme using a generally well-known music recognition method such as Mahalanobis distance, and a phoneme sequence is output. The output phoneme sequence includes ambiguities such as addition, omission, and misrecognition (phoneme recognition unit).

In S4, the CPU 26 receives the output phoneme string including the ambiguity as input, performs a dictionary search by DP matching, and outputs word candidates. In the DP matching used in this embodiment, the ambiguity of the phoneme recognition result is taken into consideration by using the similarity between phonemes as shown in FIG. Also, the inclination limit shown in FIG. 5 is used (dictionary search unit).

In S5, processing such as connection check and meaning check is performed from the extracted word candidates to output phrase candidates, and the processing ends. (Language Processing Unit) Next, the dictionary search processing in S4 will be described in detail with reference to FIG.

First, the CPU 26 reads words one by one from the dictionary stored in the ROM 28 (S11). Words are read in ascending alphabetical order as shown in FIG. Next, it is determined whether the read word is a word composed of four or less phonemes (hereinafter, a few phoneme words) or a word composed of five or more phonemes (hereinafter, a multiple phoneme word) (S12). ). If determined to be a minor phoneme word in S12, the minority phoneme word extraction unit
In S13, a word is extracted by generally well-known DP matching, and a word having a high degree of similarity is extracted. When a word is extracted, the word name and the extracted phoneme position are sent to S5. Also, one table as shown in FIG. 7 is prepared on the RAM 27 for each phoneme number word from 1 to 4 phonemes. In this table, the name of the word whose word has been extracted recently, and whether or not the word has been extracted (1, 1,
0), and the table is updated. Thereafter, the process returns to S11. Next, consider the case where a large number of phoneme words are read. The multiple phoneme word read in S11 is determined as a multiple phoneme word in S12. Next, in the majority phoneme word extraction control unit S14, it is determined whether each word is a partial word of the input majority phoneme word by referring to each phoneme number word stored in the table. If any of the words in the table is determined to be a partial word of the input word, and the word has not been extracted by the minority phoneme word extraction unit S13, the majority phoneme word is also extracted. It is determined that there is no word phoneme, and the multiple phoneme word extraction process is skipped, and the process returns to S11.
Otherwise, a multiple phoneme word extraction process is performed. In the majority phoneme word extraction unit (S15), similarly to the minority phoneme word extraction unit, word extraction is performed by generally well-known DP matching, and words having high similarity are extracted. When a word is extracted, the word name and the extracted phoneme position are sent to S5. Thereafter, the process returns to S11.

This will be described using a specific example. As an example, consider a case where the speaker utters “overseas” and the output in S3 is “kaikaini”. This output contains misrecognition in the fourth phoneme. First, in S11, the CPU 26 reads "ka" from the dictionary (see FIG. 6). Since "ka" is a word composed of two phonemes, it is determined as a minority phoneme word in S12. In the minority phoneme word extraction unit S13, the word "ka" is extracted by DP matching, and "ka" is extracted in the first and second phonemes and the fourth and fifth phonemes, and the result is sent to S5. Next, the fact that "ka" has been extracted is stored in the diphone word table (see FIG. 7), and the process returns to S11. Similarly, a minority phoneme word extraction process such as “kai” and “kaio” is performed. “Kai” was extracted in the first to third phonemes and fourth to sixth phonemes, “kaio” was not extracted, and “kai” was extracted in the three-phoneme word table (see FIG. 7). However, it is stored that “kaio” was not extracted in the 4-phoneme word table. Next, in S11, "kaigai" from the dictionary
Is read. Since “kaigai” is a word composed of six phonemes, it is determined in S12 as a multiple phoneme word. Next, the multiple phoneme word extraction control unit S14 checks whether there is a partial word of “kaigai” in the table. In the table, “ka” and “kai” “k”
It is stored that there is a partial word of "aigai", and both words are extracted. Therefore, the process is transferred to the multiple phoneme word extraction unit S15. The multiple phoneme word extraction unit S15 performs "kaigai" The word spotting of "" is performed, and "kaigai" is extracted in the first to sixth phonemes. The result is sent to S5, and then the process returns to S11.

Consider another example. Suppose that the speaker uttered "Japan" and the output of S3 was "wagakunino". CPU26 first starts with S11
Reads "ka" from the dictionary (see FIG. 6).
Since "ka" is a word composed of two phonemes,
Is determined to be a minority phoneme word. In the minority phoneme word extraction unit S13, the word "ka" is extracted by DP matching, "ka" is extracted in the third and fourth phonemes, and the result is sent to S5. Next, the fact that “ka” has been extracted is stored in the diphone word table, and the process returns to S11. Similarly, minority phoneme word extraction processing such as “kai” and “kaio” is performed, but neither word is extracted, and “kai” is not extracted in the triphoneme word table.
It is stored in the phoneme word table that “kaio” was not extracted. Next, it is assumed that “kaigai” is read from the dictionary in S11. Since “kaigai” is a word composed of six phonemes, it is determined in S12 as a multiple phoneme word. Next, in the multiple phoneme word extraction control unit S14,
It is checked whether there is a partial word "kaigai" in the table. “Ka” and “kai” in the table are “kaigai”
, And that "kai" has not been extracted. Therefore, it is determined that no word is extracted for “kaigai”, and the multiple phoneme word extraction unit S15 is skipped and the process returns to S11. Similarly, word extraction processing such as "kaigara" and "kaigi" is also skipped.

After performing this process for every word in the dictionary,
The dictionary search process ends. The present invention is not limited to the embodiments described in detail above, and various changes can be made without departing from the gist of the present invention.

For example, if there are many dropped phonemes in the phoneme recognition result output in S3, the one-phoneme word and the two-phoneme word may not be easily extracted in the minority phoneme word extraction unit S13. For this reason, in the multiple phoneme word extraction control unit S13, it may be better not to refer to the one-phoneme word and two-phoneme word tables, and it may be changed to refer to only the three-phoneme word and the four-phoneme word. is there.

Further, in the present embodiment, the minority phoneme word and the majority phoneme word are divided by four phonemes, but the invention is not limited thereto, and may be classified by other than four phonemes.

[Effects of the Invention] As is clear from the above description, in the language processing apparatus of the present invention, the word candidate searched by the search means is constituted by less than a predetermined number of phonemes by the phoneme number determination means. The first phoneme word extraction is performed on a word candidate determined to be composed of less than a predetermined number of phonemes by the phoneme number determination unit. The means performs word extraction by DP matching, and the word extraction determination means determines from the phonemes less than the predetermined number present in the word candidates determined to be composed of the predetermined number or more phonemes by the phoneme number determination means. It is determined whether or not the composed partial words have been extracted by the first phoneme word extracting means, and the partial words have been extracted by the word extraction determining means. If it is determined, the second phoneme word extraction unit extracts a word from the word candidate including the partial word by DP matching, and the word extraction determination unit does not determine that the partial word has been extracted. In this case, since the second phoneme word extracting unit does not perform DP matching on the word candidate including the partial word, the number of words for performing DP matching can be efficiently reduced, and the dictionary search can be speeded up. Becomes

[Brief description of the drawings]

1 to 7 show an embodiment embodying the present invention. FIG. 1 is a flowchart showing the flow of a dictionary search operation in the language processing apparatus of the present embodiment. FIG. FIG. 3 is a block diagram showing a configuration of the apparatus, FIG. 3 is a flowchart showing an operation flow of the language processing apparatus, and FIG.
Table showing similarity between phonemes used in matching, fifth
FIG. 6 is an explanatory diagram showing a tilt limit used in DP matching, FIG. 6 is an explanatory diagram showing a part of a dictionary, and FIG. 7 is an explanatory diagram of a table used by a dictionary search unit. In the figure, 21 is a microphone, 23 is a low-pass filter, 24 is an A / D converter, and 26 is a CPU.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-60-73596 (JP, A) JP-A-60-73596 (JP, A) JP-A-58-195895 (JP, A) JP-A-57-1995 86899 (JP, A) JP-A-63-292197 (JP, A) Furui, "Digital Speech Processing" (1985-9-25) Tokai University Press, p. 160-172 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/28 G06F 3/16 340 G10L 15/12 JICST file (JOIS)

Claims (1)

(57) [Claims] An input means for inputting linguistic information by voice, a phoneme recognizing means for analyzing the input linguistic information and performing phoneme recognition based on the analysis result, and a phoneme sequence recognized by the phoneme recognizing means. A language processing apparatus comprising: a search unit that searches for a word candidate from a phoneme-based dictionary based on the search result; and an output processing unit that outputs a sentence based on the word candidate searched by the search unit. A phoneme number determining means for determining whether the searched word candidate is a word composed of less than a predetermined number of phonemes or a word composed of a predetermined number or more of phonemes, First phoneme word extraction means for performing word extraction by DP matching on a word candidate determined to be composed of phonemes; Wherein the partial word which is composed of phonemes of less than the predetermined number present in the by-word candidate determined as being constituted by a prime first
A word extraction judging means for judging whether or not a word has been extracted by the phoneme word extracting means, and including the partial word when the word extraction judging means judges that the partial word has been extracted. A second phoneme word extracting means for extracting a word from the word candidate by DP matching, and a word including the partial word when the word extraction determining means does not determine that the partial word has been extracted. A language processing apparatus, wherein DP matching is not performed on a candidate by the second phoneme word extracting means.