JPH11143487A - Method and device for converting voice to character - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound-character conversion technology capable of converting the same word into the correct character without repetition. SOLUTION: A phoneme recognizing means 3 to divide the inputted sound into a plurality of segments and to allot one or more phonemes to each segment, a character string converting means 4 to determine one or more words based on the phoneme, and a definite word selecting means 6 to display one of the words stored in a storage part 51 on a monitor 52 as a definite word and to display other words than the fixed word as the next possible candidate, are provided.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice / character conversion technology for recognizing an input voice signal and converting it into character data.

[0002]

2. Description of the Related Art With the progress of factory automation, office automation, and home automation, many machines have been introduced in workplaces and homes. As input interfaces for transmitting commands to these machines, keyboards, mice, and the like are used. Pointing devices are mainly used. However, inputting an instruction through a keyboard is very troublesome for those who are not good at keyboard operation, and input errors are common. As a more user-friendly interface, there is a graphic interface using icons and illustrations, in which a desired command is selected from a menu displayed using a mouse or the like. In this case, it is sufficient that the number of instructions is small. However, if the number of instructions is a certain number, the hierarchy of the menu becomes deep, and much time is required to reach a desired instruction. In order to solve such disadvantages, a voice input method using voice recognition has begun to appear. The advantages of speech recognition are that it does not require skill in inputting, and that the input can be performed while other work is being performed because eyes and hands are not used.

[0003]

However, the current speech recognition technology has not been established with high reliability, and erroneous recognition often occurs. In particular, in the unspecified speaker continuous speech recognition, the same erroneous recognition continues due to the habit of a speaker or the like, and it often occurs that the same word is not recognized correctly unless the pronunciation is changed while changing the tone. An object of the present invention is to provide a voice-to-text conversion method and a voice-to-text conversion device capable of converting the same word into correct character data without repeating it many times.

[0004]

In order to achieve the above object, a speech-to-speech conversion method according to the present invention divides an input speech into a plurality of segments, recognizes phonemes, and generates one or more phonemes for each segment. And storing one or more words determined based on the phoneme in a storage unit, outputting one of the stored words as a confirmed word, and using a word other than the confirmed word as a next candidate. And

[0005] In this method, one or more words are stored in a storage unit in association with a phoneme assigned using a known phoneme recognition method, and one phoneme selected from the one or more words is stored. Is used as a control word as a control word, and a word other than the control word is set as a next display candidate. In order to check whether the confirmed word is correct, for example, it is possible to display the confirmed word on a monitor and ask the operator's judgment, or to output the confirmed word by voice and ask the operator's judgment. is there.
In any case, for example, if the first confirmed word is not correct, by sending a command for displaying some next candidate, the next candidate is output as the next confirmed word. Even if the same word is repeated many times, it is possible to avoid outputting the same wrong word repeatedly.

[0006] As a preferable method of classifying the determined word and the next display candidate, for example, a word determined based on the phoneme is stored in the storage unit with a phoneme matching probability value, and a word having a high probability value is used. It is suggested that the words be determined in order. Since the phoneme matching probability value is linked to the word corresponding to the phoneme, displaying the words in descending order of the probability value not only eliminates the error that the same character data is repeated, but also reduces the number of selections. Is more likely to encounter the correct converted character data.

Further, as a preferred embodiment of the present invention,
It is proposed to adopt a method of outputting the next candidate as a confirmed word when the same voice is input within a predetermined time, for example, displaying it on a monitor. In this case, the input of the same voice within a predetermined time is a trigger for outputting the next candidate as a fixed word instead of the fixed word first. Since different words are sequentially output as confirmed words, the operator is not given the stress that has conventionally occurred such that repeated pronunciations repeat the same erroneous conversion.

As another next candidate output trigger method, for example, a speech which is most easily recognized is set in advance as an output command of the next candidate, and if an output word is erroneously recognized, the next candidate output command is output. It is also possible to sequentially display the next candidate by pronouncing the voice as "?".

In order to achieve the above object, a speech-to-speech conversion apparatus according to the present invention divides an input speech into a plurality of segments and assigns at least one phoneme to each segment; Character string conversion means for determining one or more words based on the word, a storage unit for storing the determined word, and outputting one of the stored words as a confirmed word and displaying words other than the confirmed word And a definitive word selecting means to be the next candidate of

In this device, the character string conversion means stores one or more words in association with the assigned phonemes in the storage unit, and the confirmed word selection means outputs the selected one word as a confirmed word. In addition, a word other than the output fixed word is prepared as a next candidate of the previous fixed word while being used as a control command. When receiving a command for outputting the next candidate as a confirmed word, the next candidate is output in place of the previous confirmed word, so even if the operator repeatedly repeats the same word for correct speech recognition, Repeated display of the same incorrect word can be avoided. Of course, for checking the confirmed word, it is proposed to display the confirmed word on a monitor.

In this speech-to-speech conversion apparatus, as described in the above-mentioned proposed method, a word having a high probability value is determined by the confirmed word selecting means with respect to the word stored in the storage unit with the word matching probability value. By selecting as a confirmed word in order from, the words are output in order of the probability value, so not only the error that the same character data is repeated is eliminated, but also the correct converted character data is encountered with a small number of selections Is more likely to do so.

Further, there is provided input voice evaluation means for determining whether or not the voices input within a predetermined time are the same, and when it is determined that the same voice has been input within the predetermined time, the determined word selecting means is provided. If a configuration is adopted in which the next candidate is output as a confirmed word (for example, it is displayed on a monitor), even if the speech-to-text conversion is incorrect, different words are displayed successively by repeatedly pronouncing. The stress that the pronunciation repeats the same erroneous conversion is not given to the operator. Other features and advantages according to the present invention will become apparent from the following description of embodiments with reference to the drawings.

[0013]

FIG. 1 is a functional block diagram of a voice command input system using a voice character conversion technique according to the present invention. In this system, a surveillance camera arranged at a plant site is operated by an operator issuing a command. For example, if the operator selects "L
"NG vaporizer No. 1", the surveillance camera operates to aim at LNG vaporizer No. 1, and L is displayed on the monitor.
The state of NG vaporizer No. 1 is displayed.

This system includes a microphone 1 for converting a voice emitted by an operator into an analog voice signal, an A / D converter 2 for converting an analog voice signal sent from the microphone 1 into a digital voice signal, and a voice signal. Phoneme recognition means 3 for analyzing and replacing phonemes similar to phonetic symbols; character string conversion means 4 for assigning appropriate words to phoneme strings while accessing dictionary file 41; There is provided a storage unit 51 for temporarily storing, and a fixed word selecting unit 6 for selecting an optimum word from the words stored in the storage unit 51 as a fixed word. Further, the word selected as the confirmed word by the confirmed word selecting means 6 is displayed on the monitor 52 and is also passed to the monitoring camera control means 7. The command recognition unit 71 constituting the monitoring camera control means 7 accesses the command dictionary file 72 from the determined word passed to the monitoring camera control means 7 and derives a corresponding control command. It is converted into a control signal for operating the drive motor 74.

The phoneme recognizing means 3 uses a well-known phonemic recognition algorithm. The phoneme recognizing means 3 analyzes a temporal voice spectrum from a digital voice signal and extracts its characteristic parameters. Phoneme code conversion unit 32 for generating a phoneme code from this feature parameter
And a phoneme conversion unit 33 that gives a phoneme corresponding to the phoneme code.
And When selecting a phoneme corresponding to a phoneme code in the phoneme conversion unit 33, a plurality of phonemes are added with their phoneme matching probability values except when a phoneme is uniquely determined (that is, the phoneme matching probability value is 1). Selected. The character string conversion means 4 generates one or more words as control commands by sequentially combining the words having the highest phoneme matching probability values, and stores the words in the storage unit 51. At this time, a word matching probability value is linked to the generated word as an attribute value. The word matching probability value can be easily obtained by, for example, multiplying the phoneme matching probability value of the phoneme from which the word is based. When there are a plurality of words for a certain voice, the confirmed word selecting means 6 first displays the word having the highest word matching probability value as a confirmed word on the monitor 52 and sends it to the camera control means 7. , Start the operation of the surveillance camera.

Here, if the word displayed on the monitor 52 is different from the word pronounced by the operator, this fixed word must be canceled so that the correct word is recognized. In the case of an operator's pronunciation habit, even if the operator issues a correct control command again, the voice is not always recognized correctly. For this reason, in the present invention, words other than the confirmed word determined earlier, that is, words having the second highest word matching probability value are prepared as the next candidates in the case of misrecognition. The next candidate may be set as a confirmed word. Such replacement of the confirmed words may be performed sequentially until a correct confirmed word is output.

In this embodiment, the operator repeats the same control command again within a predetermined time (for example, 2 seconds) as a trigger for setting the next candidate as a fixed word, that is, the same voice signal is input to this system. For that purpose, an input voice evaluation means 8 is provided. The input voice evaluation unit 8 compares the voice signal input from the A / D conversion unit 2 with the voice signal input before a predetermined time, and determines a word replacement command when the two signals match each other. Output to The confirmed word selecting unit 6 that has received the word replacement command sets the next candidate word as the confirmed word, and sets the word having the next highest word matching probability as the next candidate. Of course, the voice pronounced by the operator after the elapse of the predetermined time is newly recognized.

Hereinafter, referring to FIG.
The operation of the voice command input system will be described by taking as an example the utterance of "vaporizer 1". After the voice input from the microphone 1 is gain controlled,
It is converted into a digital signal by z sampling and 16-bit quantization. Further, each frame of about 6.6 milliseconds is converted into 23 acoustic feature parameter expressions using a 20-channel filter bank or the like. This acoustic feature parameter is converted into a phoneme code by a phoneme encoder consisting of a two-stage decision tree.

In the two-stage decision tree, in the first stage of the decision tree, the acoustic features of 23 frames of the frame and the acoustic features of the frames before and after the frame are determined as the features of each frame of about 6.6 milliseconds. A total of 184 feature amounts obtained by combining 161 feature amounts derived from the relationship with the amounts are used as input vectors. The outputs of the first-stage decision tree are nine classes from No. 0 to No. 8 for performing segmentation in the next stage. These nine classes are called segment classes. Second
In the decision tree of the stage, as a feature amount, a feature amount derived from a relationship between a feature amount such as a segment class of the segment, an average of acoustic feature amounts of frames constituting the segment, and feature amounts of segments before and after the segment. Total 2
82 feature amounts are used as input vectors. The output of the second-stage decision tree is about 1800 types of phoneme codes, which are units smaller than phonemes.

The obtained phoneme code sequence is subjected to phoneme code / phoneme conversion with a phoneme matching probability with reference to a phoneme code file in which a plurality of phonemes are assigned to each phoneme code with a probability. One or more phonemes having a predetermined probability value or more are selected for each segment. This phoneme list is converted into a word while referring to the grammar. At this time, for example, the word “LPG” has a word matching probability value =
0.64 is given, and the word “LNG” is given a word matching probability value = 0.36. in this way,
The converted word is stored in the storage unit 51.

The words that have been speech-recognized and converted into characters as described above are combined by the confirmed word selecting means 6 to combine the words with the greatest probability values into the confirmed word and the next candidate in order from the most probable one. In this example, “LPG vaporizer No. 1” having the largest probability value is the first confirmed word, “LNG vaporizer No. 1” is the next candidate, and “LPG vaporizer No. 2” follows. Accordingly, “LPG vaporizer No. 1” is displayed on the monitor 52 as shown in FIG. 3, and the monitoring camera 74 starts operating to photograph the LPG vaporizer No. 1.

However, what the operator pronounced is "LNG vaporizer No. 1", and the result is erroneous recognition.
Pronounced The input voice evaluation means 8 confirms the input of the same voice signal within a predetermined period of time meaning re-input for erroneous recognition, in this case, within 2 seconds, and performs no new voice recognition processing or performs new voice recognition. Even if the processing is performed, the result is ignored, and it acts on the fixed word selecting means 6 to be the next candidate.
The LNG vaporizer No. 1 "is displayed as a confirmed word on the monitor 52, and as a result, the surveillance camera 74 changes its operation to photograph the LNG vaporizer No. 1. This time, the correct control command has been sent. Thus, if necessary, the operator will wait at least two seconds before speaking to send the next control command.

In this embodiment, the same content is reproduced once again as a trigger for displaying the next candidate.
Other methods, such as setting the most recognizable voice in advance as the display command of the next candidate, and if the displayed word is erroneously recognized, pronounces the voice as the next candidate display command. It is also possible to sequentially display the next candidates.

Further, in order to check the confirmed word, the confirmed word may be output by voice instead of being displayed on the monitor 52, so that the confirmed word may be checked by the operator.

An important point of the present invention is to use another word obtained as a result of the speech recognition performed once as a next candidate in order to avoid erroneous speech recognition in the same manner many times. It is. As long as it does not depart from the gist, it is within the scope of the present invention to use other known methods as speech recognition methods.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a voice command input system using a voice character conversion technology according to the present invention.

FIG. 2 is an explanatory diagram showing a flow of speech character conversion according to the present invention.

FIG. 3 is an explanatory diagram showing a state of a storage unit and a monitor screen in voice transcription according to the present invention;

[Explanation of symbols]

 2 A / D conversion unit 3 Phoneme recognition unit 4 Character string conversion unit 6 Fixed word selection unit 7 Camera control unit 8 Input voice evaluation unit 51 Storage unit 52 Monitor

Claims (8)

[Claims] 1. An input speech is divided into a plurality of segments, phonemes are recognized, one or more phonemes are assigned to each segment, and one or more words determined based on the phonemes are stored in a storage unit. And outputting one of the stored words as a confirmed word, and using a word other than the confirmed word as a next candidate. 2. The method according to claim 1, wherein the confirmed word is output to a monitor for display on the monitor. 3. The word determined based on the phoneme is stored in the storage unit with a phoneme matching probability value, and the words having higher probability values are determined words in order. The transcription method described in 1. 4. The voice-to-character conversion method according to claim 1, wherein if the same voice is input within a predetermined time, the next candidate is output as a confirmed word. 5. A phoneme recognition unit that divides an input speech into a plurality of segments and assigns one or more phonemes to each segment, and a character string conversion unit that determines one or more words based on the phonemes. A storage unit for storing the determined word, and a determined word selecting means for outputting one of the stored words as a determined word and selecting a word other than the determined word as a next candidate for display. Voice transcription device. 6. The apparatus according to claim 5, further comprising a monitor for displaying the confirmed word for confirming the confirmed word. 7. A word determined based on the phoneme is stored in the storage unit with a word matching probability value, and the confirmed word selecting means selects words having a higher probability value as confirmed words in order. The phonetic character conversion device according to claim 5 or 6, wherein 8. An input voice evaluation means for determining whether or not voices input within a predetermined time are the same, and when it is determined that the same voice has been input within a predetermined time, the determined word is determined. The phonetic character converter according to claim 5, wherein the selection unit outputs the next candidate as a confirmed word.