Information processing device, information processing method, and information processing program

Abstract

To realize an information processing device that can accurately correct recognition errors in speech recognition text. [Solution] The information processing device includes: acquisition means for acquiring speech recognition text converted from speech into text; error detection means for inputting the speech recognition text and a prompt for detecting an error word in speech recognition from the speech recognition text into a first large-scale language model and acquiring the error word output from the first large-scale language model; phoneme distance calculation means for acquiring one or more phoneme sequences of the pronunciation of the error word and outputting a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and sentence correction means for inputting the error word, a word correction candidate output for the error word, and a prompt instructing to select a word correction candidate to replace the error word into a second large-scale language model and reflecting the word correction candidate output from the second large-scale language model into the speech recognition text and outputting it.

Description

Technical Field

[0001] The present disclosure relates to an information processing apparatus, an information processing method, and an information processing program.

Background Art

[0002] There is known a speech recognition technology for mechanically generating text from voice data obtained by recording human speech. As an example of such a technology, for example, the speech recognition technology described in Non-Patent Document 1 can be cited.

Prior Art Documents

Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Non-Patent Document 1 discloses a speech recognition correction technology for detecting recognition errors in speech recognition text obtained by converting voice data into text, generating correction candidates for the recognition errors, and selecting the most appropriate one from the correction candidates. However, in the technology of Non-Patent Document 1, since correction candidates are generated according to the context, when the voice data has particularly highly specialized content, it may not be possible to appropriately generate correction candidates for recognition errors. That is, in a specialized field, there is a risk that the correction accuracy will not be improved much.

[0005] The present disclosure has been made in view of the above problems, and an exemplary object thereof is to provide a technology for accurately correcting recognition errors in speech recognition text.

Means for Solving the Problems

[0006] An information processing device relating to an exemplary aspect of this disclosure includes: acquisition means for acquiring speech recognition text converted from speech into text; error detection means for inputting the speech recognition text and a prompt for detecting an error word in speech recognition from the speech recognition text into a first large-scale language model and acquiring the error word output from the first large-scale language model; phoneme distance calculation means for acquiring one or more phoneme sequences of the reading of the error word and outputting a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and sentence correction means for inputting the error word, the word correction candidate output for the error word, and a prompt for instructing to select the word correction candidate to replace the error word into a second large-scale language model and reflecting the word correction candidate output from the second large-scale language model into the speech recognition text and outputting it.

[0007] An information processing method relating to an exemplary aspect of this disclosure includes: obtaining speech recognition text converted from speech into text; inputting the speech recognition text and a prompt for detecting an error word in speech recognition from the speech recognition text into a first large-scale language model, and obtaining the error word output from the first large-scale language model; obtaining one or more phoneme sequences of the reading of the error word, and outputting a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; inputting the error word, the word correction candidate output for the error word, and a prompt instructing the system to select a word correction candidate to replace the error word into a second large-scale language model, and reflecting the word correction candidate output from the second large-scale language model into the speech recognition text and outputting it.

[0008] An information processing program relating to an exemplary aspect of this disclosure causes a computer to perform the following steps: a process to obtain speech recognition text converted from speech into text; a process to input the speech recognition text and a prompt to detect an error word in speech recognition from the speech recognition text into a first large-scale language model, and obtain the error word output from the first large-scale language model; a process to obtain one or more phoneme sequences of the pronunciation of the error word and output a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; a process to input the error word, the word correction candidate output for the error word, and a prompt instructing the computer to select the word correction candidate to replace the error word into a second large-scale language model, and output the word correction candidate output from the second large-scale language model in accordance with the speech recognition text. [Effects of the Invention]

[0009] One exemplary aspect of this disclosure is that it can provide a technology that accurately corrects recognition errors in speech recognition text. [Brief explanation of the drawing]

[0010] [Figure 1] This is a block diagram showing the configuration of the information processing device related to this disclosure. [Figure 2] This is a flowchart showing the flow of the information processing method related to this disclosure. [Figure 3] This is a block diagram showing the configuration of the information processing device related to this disclosure. [Figure 4] This is a schematic diagram showing an example of a word pronunciation dictionary related to this disclosure. [Figure 5] This is a schematic diagram showing an example of a phoneme distance table related to this disclosure. [Figure 6] A flowchart illustrating an example of information processing performed by the information processing device 1A related to this disclosure. [Figure 7] This is a schematic diagram illustrating a method for generating a phoneme distance table using the trained machine model described herein. [Figure 8]This is a block diagram showing the configuration of a computer that functions as an information processing device related to this disclosure. [Modes for carrying out the invention]

[0011] The following are examples of embodiments of the present invention. However, the present invention is not limited to the exemplary embodiments shown below, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining some or all of the technologies (things or methods) employed in each of the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, embodiments obtained by appropriately omitting some of the technologies employed in each of the exemplary embodiments shown below may also be included in the scope of the present invention. In addition, the effects mentioned in each of the exemplary embodiments shown below are examples of effects that can be expected in that exemplary embodiment and do not define the scope of the present invention. That is, embodiments that do not produce the effects mentioned in each of the exemplary embodiments shown below may also be included in the scope of the present invention.

[0012] [First Exemplary Embodiment]

[0013] A first exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is the basic form for each of the exemplary embodiments described later. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur. Furthermore, each technology shown in the drawings referenced to explain this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur.

[0014] (Configuration of information processing device) The configuration of the information processing device 1 will be explained with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the information processing device 1. The information processing device 1 is a device that detects recognition errors in speech recognition text converted from speech data and outputs the correct speech recognition text. As shown in Figure 1, the information processing device 1 comprises an acquisition unit 11 (acquisition means as defined in the claims), an error detection unit 12 (error detection means as defined in the claims), a phoneme distance calculation unit 13 (phoneme distance calculation means as defined in the claims), and a text correction unit 14 (text correction means as defined in the claims). Each part of the information processing device 1 will be described below.

[0015] The acquisition unit 11 acquires speech recognition text, which is a transcription of speech into text. Speech recognition text can be generated from recorded speech data using known techniques. Speech recognition text (hereinafter also simply referred to as "text") may be recorded in any memory or database, and the acquisition unit 11 may acquire pre-recorded speech recognition text and record it in the memory of the information processing device 1. Alternatively, the acquisition unit 11 may use a program that generates speech recognition text from speech data to generate speech recognition text from speech data, record it in the memory of the information processing device 1, and then acquire it.

[0016] The error detection unit 12 inputs the speech recognition text and a prompt to detect incorrect words in the speech recognition text to the first large-scale language model, and acquires the incorrect words output from the first large-scale language model. The large-scale language model (LLM) is any existing neural network model that has been trained using a large amount of language data. The error detection unit 12 inputs a prompt to this large-scale language model, along with the speech recognition text, such as "Please extract the incorrect words from the following sentence," and the large-scale language model outputs incorrect words (words that are considered incorrect) from places where the context does not connect. The error detection unit 12 acquires these outputted incorrect words.

[0017] The phoneme distance calculation unit 13 acquires one or more phoneme sequences of the pronunciation of the misspelled word, and outputs a word correction candidate whose normalized phoneme distance between two phonemes and the phoneme sequence is below a predetermined threshold value. When the speech recognition text is in Japanese, the misspelled word may be not only a word (Chinese characters, hiragana, katakana, etc.), but also a katakana sequence or a character string that is not a word. When the speech recognition text is in English, the misspelled word is a group of alphabetic words. The phoneme distance calculation unit 13 acquires one or more phoneme sequences of the pronunciation of such misspelled words. When there are multiple pronunciations in the case of Chinese characters, the phoneme distance calculation unit 13 acquires multiple pronunciations. Then, the phoneme distance calculation unit 13 outputs a word correction candidate consisting of a phoneme sequence whose normalized phoneme distance between two phonemes and the acquired phoneme sequence is below a predetermined threshold value. A phoneme is the smallest unit of sound corresponding to a consonant or a vowel. Therefore, it is not the same as a syllable. For example, the vowel phonemes are a, i, u, e, o, and the consonant phonemes are k (the ka row), s (the sa row), t (the ta row), etc. Also included are nasal sounds, stop sounds, etc. Punctuation marks may be considered as silent phonemes. The phoneme distance is an index representing the ease of recognizing the difference between two phonemes. For example, the larger the phoneme distance, the greater the difference, that is, it is judged that the phoneme is less likely to be misspelled. Therefore, a word correction candidate consisting of a phoneme sequence whose total normalized (normalized) phoneme distance is below a predetermined threshold value is selected and output. "Normalization" means, for example, dividing the total value of the phoneme distance by the number of phonemes. Since a word is composed of multiple phonemes, as the length of the word increases, the total phoneme distance also increases. Therefore, by dividing the total value of the phoneme distance by the number of phonemes, a phoneme distance comparable between words can be obtained.

[0018] The text correction unit 14 inputs an incorrect word, a word correction candidate output for the incorrect word, and a prompt instructing to select a word correction candidate to replace the incorrect word into a second large language model, and reflects the word correction candidate output from the second large language model in the speech recognition text and outputs it. The text correction unit 14 generates and inputs a prompt such as "Please select a word to correct the incorrect word in the text from the word correction candidates" to the second large language model, together with the speech recognition text, the incorrect word, and the word correction candidate output for the incorrect word. Then, the text correction unit 14 acquires the selected word output from the second large language model. Further, the text correction unit 14 generates and outputs a corrected speech recognition text by replacing it with the selected word. Note that the first large language model and the second large language model may be the same large language model.

[0019] Alternatively, the text correction unit 14 may generate a prompt such as "Please create the correct text by replacing the incorrect word in the text with the most suitable word correction candidate" instead of the above prompt, input it into the second large language model, acquire the entire text of the output "correct text", and output it as the corrected text as it is.

[0020] (Effect of the information processing apparatus 1) As described above, the information processing device 1 employs a configuration comprising: an acquisition unit that acquires speech recognition text converted from speech into text; an error detection unit that inputs the speech recognition text and a prompt to detect erroneous words in speech recognition from the speech recognition text into a first large-scale language model and acquires the erroneous words output from the first large-scale language model; a phoneme distance calculation unit that acquires one or more phoneme sequences of the pronunciation of the erroneous word and outputs word correction candidates for which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and a text correction unit that inputs the erroneous word, the word correction candidates output for the erroneous word, and a prompt instructing the user to select a word correction candidate to replace the erroneous word into a second large-scale language model and outputs the result output from the second large-scale language model in the speech recognition text. Therefore, the information processing device 1 has the effect of correcting recognition errors in speech recognition text with higher accuracy than conventional technology by analyzing the phonemes of the erroneous words.

[0021] (Information processing flow) The flow of the information processing method S1 will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the information processing method S1. As shown in Figure 2, the information processing method S1 includes a text acquisition process S11, an error word acquisition process S12, a word correction candidate output process S13, and a corrected text output process S14.

[0022] Text acquisition process S11 is the process of acquiring speech recognition text, which is the text generated from speech. Text acquisition process S11 is executed by the acquisition unit 11 (one processor). The contents of text acquisition process S11 are as described in the section on the acquisition unit 11.

[0023] The error word acquisition process S12 is a process that inputs the speech recognition text and a prompt to detect error words in the speech recognition from the speech recognition text into the first large-scale language model, and acquires the error words output from the first large-scale language model. The error word acquisition process S12 is executed by the error detection unit 12 (one processor). The contents of the error word acquisition process S12 are as described with respect to the error detection unit 12.

[0024] The word correction candidate output process S13 is a process that obtains one or more phoneme sequences of the pronunciation of the incorrect word and outputs word correction candidates for which the normalized phoneme distance between the two phonemes in the phoneme sequence is less than or equal to a predetermined threshold. The word correction candidate output process S13 is executed by the phoneme distance calculation unit 13 (one processor). The contents of the word correction candidate output process S13 are as described with respect to the phoneme distance calculation unit 13.

[0025] The corrected text output process S14 inputs the erroneous word, the word correction candidate output for that erroneous word, and a prompt instructing the system to select a word correction candidate to replace the erroneous word into the second large-scale language model, and then reflects the word correction candidate output from the second large-scale language model into the speech recognition text and outputs it. The corrected text output process S14 is executed by the text correction unit 14 (one processor). The contents of the corrected text output process S14 are as described with respect to the text correction unit 14.

[0026] (Effects of information processing methods) As described above, the information processing method S1 employs a configuration in which at least one processor performs the following: a text acquisition process to acquire speech recognition text converted from speech; an error word acquisition process that inputs the speech recognition text and a prompt to detect erroneous words in speech recognition from the speech recognition text to a first large-scale language model and acquires the erroneous words output from the first large-scale language model; a word correction candidate output process that acquires one or more phoneme sequences of the pronunciation of the erroneous word and outputs word correction candidates in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and a corrected text output process that inputs the erroneous word, the word correction candidates output for the erroneous word, and a prompt instructing the system to select a word correction candidate to replace the erroneous word to a second large-scale language model and reflects the word correction candidates output from the second large-scale language model in the speech recognition text and outputs it. Therefore, according to the information processing method S1, recognition errors in speech recognition text can be corrected with higher accuracy than conventional techniques by analyzing the phonemes of the erroneous words.

[0027] [Second exemplary embodiment] A second exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiment are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise.

[0028] (Configuration of Information Processing Device 1A) The configuration of the information processing device 1A will be described with reference to Figure 3. Figure 3 is a block diagram showing the configuration of the information processing device 1A. In addition to the acquisition unit 11, error detection unit 12, phoneme distance calculation unit 13, and text correction unit 14 of the information processing device 1, the information processing device 1A includes an input / output interface (input / output IF) 20, at least one processor 30, and at least one memory 40. The phoneme distance calculation unit 13 also includes a word reading dictionary 131 and a phoneme distance table 132. The information processing device 1A may also be connected to a display unit (display) 70. The functions of each part of the information processing device 1A other than those of the information processing device 1 described in Exemplary Embodiment 1 will be described below.

[0029] The processor 30 can be configured using at least one general-purpose processor such as an MPU (Micro Processing Unit) or CPU (Central Processing Unit). The processor 30 may also include a dedicated processor composed of an ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Gate Array), or PLD (Programmable Logic Device).

[0030] Memory 40 may include multiple types of memory, such as ROM (Read Only Memory) and RAM (Random Access Memory). Memory 40 may also include internal or external memory such as an HDD (Hard Disk Drive) or SSD (Solid State Drive). For example, the processor 20 performs the functions of an acquisition unit 11, an error detection unit 12, a phoneme distance calculation unit 13, and a text correction unit 14 by loading various control programs recorded in the ROM of memory 40 into the RAM and executing them. Furthermore, data such as various programs and speech recognition text may be recorded in an externally located cloud database (not shown).

[0031] The input / output IF20 is an interface for sending and receiving data to and from the outside. Communication between the input / output IF20 and the outside may be performed, for example, via the Internet 100. The input / output IF20 may be equipped with a short-range communication device such as WiFi® or Bluetooth® that can connect wirelessly to an internet connection point. Alternatively, it may be a wired connection interface such as a USB connector. For example, communication between the first large-scale language model 50 and the second large-scale language model 60 is performed via the Internet 100.

[0032] The error detection unit 12 may acquire information about the topic of the speech recognition text along with the erroneous word. This topic information may be a concept that indicates the topic, or it may be words that frequently appear within that topic. The phoneme distance calculation unit 13 can use this topic information to narrow down the word correction candidates. For example, if a large number of word correction candidates are listed, the phoneme distance calculation unit 13 may evaluate how relevant each word correction candidate is to the topic and extract those with high relevance.

[0033] The phoneme distance calculation unit 13 obtains the phoneme sequence of the incorrect word using a word pronunciation dictionary. An example of a word pronunciation dictionary is shown in Figure 4. The word pronunciation dictionary 131 shown in Figure 4 is a dictionary that associates words (kanji) with their pronunciations (or phoneme sequences). For example, the word pronunciation dictionary 131 records that the pronunciation (phoneme sequence in hiragana) of the word "motive" is "dōki," and that the phoneme sequence representing it in the alphabet is "douki." The same pronunciation (phoneme sequence) is also recorded for the words "palpitations" and "synchronization." Note that the word pronunciation dictionary 131 may contain only the word and one of the phoneme sequences.

[0034] The phoneme distance calculation unit 13 may derive the phoneme distance using a phoneme distance table in which the distance between two phonemes is defined. An example of a phoneme distance table is shown in Figure 5. Phoneme distance table 132 shows a table that records the phoneme distances between phonemes in the "a" row (a, i, u, e, o). For example, the phonemes "a" and "a" are the same, so the phoneme distance is zero. The phoneme distance between "a" and "i" is 0.9. The phoneme distance is a numerical value between zero and one, and the smaller the phoneme distance (the closer the pronunciation), the smaller the value. There is also a table that records the phoneme distance between phonemes in the "a" row and phonemes in other rows, and similar tables exist for phonemes other than those in the "a" row.

[0035] The phoneme distance calculation unit 13 outputs word correction candidates by selecting and outputting word correction candidates that yield the smallest possible difference from the incorrect word. In other words, the phoneme distance table 132 is a cost table, and words with combinations of phonemes that result in the smallest possible cost calculated using this table are selected as word correction candidates. The phoneme distance table 132 is created in advance.

[0036] The phoneme distance table 132 may be, for example, an inter-phoneme cost table created by a model trained using paired data of incorrect words and corresponding correct words contained in the recognized text of speech acquired under common conditions. Common conditions refer to conditions where the topic (domain), the recording environment of the speech data (location, room, microphone, etc.), and the speech recognition model are similar or the same. A machine model is trained using a large amount of training data containing incorrect words and correct words contained in the recognized text of speech data acquired under such conditions. The phoneme distance (cost) of any two phonemes can then be evaluated using this trained machine model and compiled into a table.

[0037] Figure 7 is a schematic diagram illustrating a method for generating a phoneme distance table using a trained machine model. First, a dataset containing an incorrect word 1A and a correct word 1B is defined as pair D1. Training data D, containing n such pairs, is input into an untrained machine model M for training. This training process is repeated to generate a trained machine model LM. The machine model M can learn the likelihood of error (cost) between phonemes from the combination of phoneme sequences of the incorrect and correct words. The phoneme distance table output from the trained machine model LM can then be used as the phoneme distance table 132 by the phoneme distance calculation unit 13.

[0038] If the phoneme distance calculation unit 13 finds no word correction candidates whose normalized phoneme distance is less than or equal to a predetermined threshold, it may output the word correction candidate with the smallest normalized phoneme distance among the evaluated word correction candidates. Alternatively, it may output multiple word correction candidates, including the word correction candidate with the smallest normalized phoneme distance.

[0039] The word-reading dictionary may have flags indicating specialized terminology. These flags may be assigned by the user (expert), or they may be assigned using, for example, a list of specialized terms in the target field collected in advance by the phoneme distance calculation unit 13 using an LLM or a publicly available specialized terminology dictionary. In the word-reading dictionary 131 shown in Figure 4, the words "palpitations" and "tumor" are flagged with the specialized term flag TA. This word-reading dictionary 131 is a dictionary for correcting errors in speech recognition text in the medical field. Therefore, "palpitations" and "tumor" are flagged with TA as medical specialized terms. The phoneme distance calculation unit 13 may add weights to words with the flag TA when calculating the phoneme distance. Adding weights is a process that increases the evaluation value, and in this exemplary embodiment, it corresponds to a process that reduces the cost.

[0040] The text correction unit 14 may use the second large-scale language model generated by the search extension to select word correction candidates based on the information acquired by the error detection unit 12. Search extension generation (RAG) is a method for accurately correcting errors in speech recognition text related to a particular field by, for example, inputting data on specialized terminology into a large-scale language model and retraining it.

[0041] Specifically, the error detection unit 12 performs error detection using, for example, a general-purpose large-scale language model tuned for medical use. It then narrows down the text area from the remaining words that were not identified as errors. If the error detection unit 12 is able to narrow down the text content to, for example, a medical department, it sends that information to the text correction unit 14. The text correction unit 14 uses a second large-scale language model, which has been generated by narrowing the search to the "medical department" area, to select word correction candidates. In this way, errors in speech recognition text can be corrected with high accuracy.

[0042] Figure 6 is a flowchart showing an example of an information processing method S2 executed by the information processing device 1A. First, the acquisition unit 11 acquires speech recognition text TX (step S21). Suppose there is a sentence in it that says, "Palpitations and dizziness occur due to anemia or low blood pressure." In response, the error detection unit 12 acquires the word "motive" as an error word (step S22). Next, the phoneme distance calculation unit 13 acquires the phoneme sequence for the reading of "motive". For example, the phoneme distance calculation unit 13 acquires "doki" using the word reading dictionary 131 shown in Figure 4. Next, the phoneme distance calculation unit 13 outputs word correction candidates for which the normalized phoneme distance between the two phonemes of the phoneme sequence "doki" is less than or equal to a predetermined threshold. For example, it outputs three words that have the same reading but a total phoneme distance of zero: "motive", "palpitations", and "synchronization" (step S23). Next, the sentence correction unit 14 inputs these words together with the speech recognition text into the second large-scale language model. Then, the correct text output from the second large-scale language model, "Palpitations and dizziness occur due to anemia or low blood pressure," is obtained and replaced with the corresponding sentence in the original speech recognition text (step S24).

[0043] (Effects of Information Processing Device 1A) As described above, in the information processing device 1A, the phoneme distance calculation unit 13 employs a configuration that uses a word pronunciation dictionary to obtain the phoneme sequence of the erroneous word. Therefore, in addition to the effects achieved by the information processing device 1, the information processing device 1A also has the effect of efficiently obtaining the correct phoneme sequence of the erroneous word.

[0044] Furthermore, in the information processing device 1A, the phoneme distance calculation unit 13 employs a configuration in which the phoneme distance is derived using a phoneme distance table in which the distance between two phonemes is defined. Therefore, in addition to the effects of the information processing device 1, the information processing device 1A also has the effect of efficiently deriving the phoneme distance.

[0045] Furthermore, in the information processing device 1A, the error detection unit 12 acquires information about the topic of the speech recognition text along with the erroneous word, and the phoneme distance calculation unit 13 uses this information to narrow down the candidate words for correction. Therefore, in addition to the effects achieved by the information processing device 1, the information processing device 1A also has the effect of being able to narrow down the candidate words for correction with high accuracy.

[0046] Furthermore, in the information processing device 1A, the text correction unit 14 is configured to use the second large-scale language model generated by the search extension to select word correction candidates based on the information acquired by the error detection unit 12. Therefore, in addition to the effects achieved by the information processing device 1, the information processing device 1A also has the effect of being able to select word correction candidates with greater accuracy.

[0047] Furthermore, in the information processing device 1A, the phoneme distance calculation unit 13 is configured to output the word correction candidate with the smallest normalized phoneme distance if there are no word correction candidates whose normalized phoneme distance is below a predetermined threshold. Therefore, in addition to the effects achieved by the information processing device 1, the information processing device 1A can output the most appropriate word correction candidate even if no word correction candidate that satisfies the predetermined conditions is found.

[0048] Furthermore, in the information processing device 1A, the word reading dictionary 131 is flagged for specialized terminology, and the phoneme distance calculation unit 13 is configured to add weights to words with this flag when calculating the phoneme distance. Therefore, in addition to the effects of the information processing device 1, the information processing device 1A can perform speech recognition text correction with greater accuracy for predetermined specialized fields.

[0049] Furthermore, in the information processing device 1A, the phoneme distance table 132 is configured to be an inter-phoneme cost table created by a model trained using pair data of incorrect words and corresponding correct words contained in the recognized speech text acquired under common conditions. Therefore, in addition to the effects of the information processing device 1, the information processing device 1A provides the effect of being able to correct the recognized speech text with greater accuracy for speech recognized text acquired under specific conditions.

[0050] [Examples of implementation using software] Some or all of the functions of the information processing devices 1,1A (hereinafter also referred to as "each of the above devices") may be implemented by hardware such as integrated circuits (IC chips) or by software.

[0051] In the latter case, each of the above devices is implemented, for example, by a computer that executes instructions for a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as Computer C) is shown in Figure 8. Figure 8 is a block diagram showing the hardware configuration of Computer C, which functions as each of the above devices.

[0052] Computer C comprises at least one processor C1 and at least one memory C2. Memory C2 stores a program P that causes computer C to operate as each of the above-mentioned devices. In computer C, processor C1 reads program P from memory C2 and executes it, thereby realizing each of the above-mentioned devices.

[0053] For processor C1, for example, a CPU (Central Processing Unit), GPU (Graphic Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof can be used. For memory C2, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination thereof can be used.

[0054] Computer C may also be equipped with RAM (Random Access Memory) for loading program P at runtime and for temporarily storing various data. Furthermore, computer C may be equipped with communication interfaces for sending and receiving data with other devices. Additionally, computer C may be equipped with input / output interfaces for connecting input / output devices such as keyboards, mice, displays, and printers.

[0055] Furthermore, program P can be recorded on a non-temporary, tangible recording medium M that is readable by computer C. Such a recording medium M could be, for example, tape, disk, card, semiconductor memory, or programmable logic circuitry. Computer C can acquire program P via such a recording medium M. Program P can also be transmitted via a transmission medium. Such a transmission medium could be, for example, a communication network or broadcast waves. Computer C can also acquire program P via such a transmission medium.

[0056] Furthermore, each of the above functions of each of the above devices may be implemented by a single processor in a single computer, by multiple processors in a single computer working together, or by multiple processors in each of multiple computers working together. In addition, the programs for implementing each of the above functions in each of the above devices may be stored in a single memory in a single computer, distributed and stored in multiple memories in a single computer, or distributed and stored in multiple memories in each of multiple computers.

[0057] [Additional Note 1] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0058] (Note 1) Information processing device comprising: acquisition means for acquiring speech recognition text converted from speech into text; error detection means for inputting the speech recognition text and a prompt for detecting an error word in speech recognition from the speech recognition text into a first large-scale language model and acquiring the error word output from the first large-scale language model; phoneme distance calculation means for acquiring one or more phoneme sequences of the pronunciation of the error word and outputting a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and text correction means for inputting the error word, the word correction candidate output for the error word, and a prompt for instructing to select the word correction candidate to replace the error word into a second large-scale language model and reflecting the word correction candidate output from the second large-scale language model into the speech recognition text and outputting it.

[0059] (Note 2) The phoneme distance calculation means is an information processing device according to Appendix 1, which obtains the phoneme sequence of the error word using a word pronunciation dictionary.

[0060] (Note 3) The phoneme distance calculation means derives the phoneme distance using a phoneme distance table in which the distance between two phonemes is defined, as described in Appendix 1 or 2.

[0061] (Note 4) The information processing apparatus according to any one of the appendices 1 to 3, wherein the error detection means acquires information about the topic of the speech recognition text along with the error word, and the phoneme distance calculation means uses the information to narrow down the candidate words for correction.

[0062] (Note 5) The information processing device described in Appendix 4, wherein the text correction means uses a second large-scale language model generated by search extension to cause the error detection means to select word correction candidates based on the information acquired.

[0063] (Note 6) The phoneme distance calculation means outputs the word correction candidate with the smallest normalized phoneme distance if there are no word correction candidates whose normalized phoneme distance is less than or equal to a predetermined threshold, as described in any one of appendices 1 to 5.

[0064] (Note 7) The information processing device described in Appendix 2, wherein the word pronunciation dictionary is flagged for technical terms, and the phoneme distance calculation means adds weights to words that are flagged when calculating the phoneme distance.

[0065] (Note 8) The information processing device described in Appendix 3, wherein the phoneme distance table is an inter-phoneme cost table created by a model trained using paired data of incorrect words and corresponding correct words contained in the recognized text of speech acquired under common conditions.

[0066] (Note 9) An information processing method comprising: obtaining speech recognition text converted from speech into text; inputting the speech recognition text and a prompt for detecting an erroneous word in speech recognition from the speech recognition text into a first large-scale language model and obtaining the erroneous word output from the first large-scale language model; obtaining one or more phoneme sequences for the reading of the erroneous word and outputting a word correction candidate for which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; inputting the erroneous word, the word correction candidate output for the erroneous word, and a prompt instructing the system to select a word correction candidate to replace the erroneous word into a second large-scale language model and reflecting the word correction candidate output from the second large-scale language model into the speech recognition text and outputting it.

[0067] (Note 10) An information processing program that causes a computer to perform the following steps: a process to obtain speech recognition text converted from speech into text; inputting the speech recognition text and a prompt to detect erroneous words in speech recognition from the speech recognition text into a first large-scale language model; a process to obtain the erroneous words output from the first large-scale language model; a process to obtain one or more phoneme sequences of the pronunciation of the erroneous words and output word correction candidates for which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; inputting the erroneous words, the word correction candidates output for the erroneous words, and a prompt instructing the computer to select a word correction candidate to replace the erroneous words into a second large-scale language model; and a process to reflect the word correction candidates output from the second large-scale language model into the speech recognition text and output it.

[0068] [Additional Note 2] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0069] (Note 21) Information processing device comprising at least one processor, the at least one processor performing: an acquisition process to acquire speech recognition text converted from speech into text; an error detection process that inputs the speech recognition text and a prompt to detect an error word in speech recognition from the speech recognition text to a first large-scale language model and acquires the error word output from the first large-scale language model; a phoneme distance calculation process that acquires one or more phoneme sequences of the reading of the error word and outputs a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and a sentence correction process that inputs the error word, the word correction candidate output for the error word, and a prompt instructing to select the word correction candidate to replace the error word to a second large-scale language model and reflects the word correction candidate output from the second large-scale language model in the speech recognition text and outputs it.

[0070] The information processing device may also include memory. Furthermore, the memory may store a program that causes at least one processor to execute each of the aforementioned processes.

[0071] (Note 22) The information processing apparatus according to Appendix 21, wherein in the phoneme distance calculation process, at least one processor obtains the phoneme sequence of the error word using a word pronunciation dictionary.

[0072] (Note 23) The information processing apparatus according to Appendix 21, wherein in the phoneme distance calculation process, at least one processor derives the phoneme distance using a phoneme distance table in which the distance between two phonemes is defined.

[0073] (Note 24) The information processing apparatus according to Appendix 21, wherein in the error detection process, the at least one processor obtains information about the topic of the speech recognition text along with the error word, and the phoneme distance calculation process uses this information to narrow down the candidate words for correction.

[0074] (Note 25) The information processing apparatus according to Appendix 24, wherein in the text correction process, the at least one processor uses a second large-scale language model generated by search extension to select word correction candidates based on the information acquired by the error detection process.

[0075] (Note 26) The information processing apparatus according to Appendix 21, wherein, in the phoneme distance calculation process, if there are no word correction candidates whose normalized phoneme distance is less than or equal to a predetermined threshold, the at least one processor outputs the word correction candidate with the smallest normalized phoneme distance.

[0076] (Note 27) The aforementioned word-reading dictionary is flagged for technical terms, and the phoneme distance calculation process adds weights to words that are flagged when calculating the phoneme distance, as described in Appendix 22.

[0077] (Note 28) The information processing device described in Appendix 23, wherein the phoneme distance table is an inter-phoneme cost table created by a model trained using paired data of incorrect words and corresponding correct words contained in the recognized text of speech acquired under common conditions.

[0078] (Note 29) An information processing method comprising: an acquisition process in which at least one processor acquires speech recognition text converted from speech into text; an error detection process in which the speech recognition text and a prompt for detecting an error word in speech recognition are input to a first large-scale language model and the error word output from the first large-scale language model is acquired; a phoneme distance calculation process in which one or more phoneme sequences of the pronunciation of the error word are acquired and word correction candidates are output for which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and a sentence correction process in which the error word, the word correction candidates output for the error word, and a prompt instructing the system to select a word correction candidate to replace the error word are input to a second large-scale language model and the word correction candidates output from the second large-scale language model are reflected in the speech recognition text and output.

[0079] [Additional Note 3] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0080] (Note 31) Information processing method comprising: an acquisition process in which at least one processor acquires speech recognition text converted from speech into text; an error detection process in which the at least one processor inputs the speech recognition text and a prompt to a first large-scale language model to detect an error word in speech recognition from the speech recognition text, and acquires the error word output from the first large-scale language model; a phoneme distance calculation process in which the at least one processor acquires one or more phoneme sequences of the pronunciation of the error word and outputs a word correction candidate in which the normalized phoneme distance between two phonemes with the phoneme sequence is less than or equal to a predetermined threshold; and a sentence correction process in which the at least one processor inputs the error word, the word correction candidate output for the error word, and a prompt to instruct the second large-scale language model to select a word correction candidate to replace the error word, and reflects the word correction candidate output from the second large-scale language model in the speech recognition text and outputs it.

[0081] (Note 32) The information processing method according to Appendix 31, wherein in the phoneme distance calculation process, at least one processor obtains the phoneme sequence of the error word using a word pronunciation dictionary.

[0082] (Note 33) The information processing method according to Appendix 31 or 32, wherein in the phoneme distance calculation process, at least one processor derives the phoneme distance using a phoneme distance table in which the distance between two phonemes is defined.

[0083] (Note 34) The information processing method according to any one of appendices 31 to 33, wherein in the error detection process, the at least one processor obtains information about the topic of the speech recognition text along with the error word, and in the phoneme distance calculation process, uses this information to narrow down the candidate words for correction.

[0084] (Note 35) The information processing method according to Appendix 34, wherein in the text correction process, the at least one processor uses the second large-scale language model generated by the search extension to select word correction candidates based on the acquired information.

[0085] (Note 36) The information processing method according to any one of appendices 31 to 35, wherein in the phoneme distance calculation process, if there are no word correction candidates whose normalized phoneme distance is less than or equal to a predetermined threshold, the at least one processor outputs the word correction candidate with the smallest normalized phoneme distance.

[0086] (Note 37) The aforementioned word-reading dictionary is flagged for technical terms, and the information processing method described in Appendix 32 adds weights to words that are flagged in the phoneme distance calculation process when determining the phoneme distance.

[0087] (Note 38) The information processing method described in Appendix 33, wherein the phoneme distance table is an inter-phoneme cost table created by a model trained using paired data of incorrect words and corresponding correct words contained in the recognized text of speech acquired under common conditions. [Explanation of symbols]

[0088] 1,1A...Information Processing Device 11...Acquisition part 12. Error detection unit 13... Phoneme distance calculation section 131...Word Pronunciation Dictionary 132... Phoneme distance table 14...Sentence correction department 20. Input / Output Interface 31... Processor 40...memory 50. The first large-scale language model 60...Second large-scale language model 70...Display section 100...Internet

Claims

1. A means for obtaining speech recognition text, which is a transcription of speech into text, An error detection means inputs the speech recognition text and a prompt for detecting an error word in the speech recognition from the speech recognition text into a first large-scale language model, and obtains the error word output from the first large-scale language model. A phoneme distance calculation means that obtains one or more phoneme sequences of the pronunciation of the aforementioned erroneous word and outputs a word correction candidate in which the normalized phoneme distance between two phonemes in the aforementioned phoneme sequence is less than or equal to a predetermined threshold, A text correction means inputs the aforementioned erroneous word, the word correction candidate output for the erroneous word, and a prompt instructing the user to select the word correction candidate to replace the erroneous word, and reflects the word correction candidate output from the second large-scale language model into the speech recognition text and outputs it. An information processing device equipped with the following features.

2. The information processing apparatus according to claim 1, wherein the phoneme distance calculation means obtains the phoneme sequence of the error word using a word pronunciation dictionary.

3. The information processing apparatus according to claim 1, wherein the phoneme distance calculation means derives the phoneme distance using a phoneme distance table in which the distance between two phonemes is defined.

4. The information processing apparatus according to any one of claims 1 to 3, wherein the error detection means acquires information about the topic of the speech recognition text along with the error word, and the phoneme distance calculation means uses the information to narrow down the candidate words for correction.

5. The information processing apparatus according to claim 4, wherein the text correction means uses a second large-scale language model generated by search extension to cause the error detection means to select word correction candidates based on the information acquired.

6. The information processing apparatus according to any one of claims 1 to 3, wherein the phoneme distance calculation means outputs the word correction candidate with the smallest normalized phoneme distance if there are no word correction candidates whose normalized phoneme distance is less than or equal to a predetermined threshold.

7. The information processing apparatus according to claim 2, wherein the word pronunciation dictionary is assigned a flag for technical terms, and the phoneme distance calculation means adds a weight to the words that are assigned the flag when calculating the phoneme distance.

8. The information processing apparatus according to claim 3, wherein the phoneme distance table is an inter-phoneme cost table created by a model trained using paired data of incorrect words and corresponding correct words contained in the recognized text of speech acquired under common conditions.

9. Obtaining speech recognition text by converting speech into text, The process involves inputting the aforementioned speech recognition text and a prompt for detecting erroneous words in the speech recognition text into a first large-scale language model, and obtaining the erroneous words output from the first large-scale language model. The system obtains one or more phoneme sequences of the pronunciation of the aforementioned erroneous word, and outputs a word correction candidate in which the normalized phoneme distance between two phonemes in the aforementioned phoneme sequence is less than or equal to a predetermined threshold. The erroneous word, the word correction candidate output for the erroneous word, and a prompt instructing the user to select the word correction candidate to replace the erroneous word are input to a second large-scale language model, and the word correction candidate output from the second large-scale language model is reflected in the speech recognition text and output. Information processing methods, including those mentioned above.

10. On the computer, The process of obtaining speech recognition text by converting speech into text, The process involves inputting the speech recognition text and a prompt for detecting erroneous words in the speech recognition text into a first large-scale language model, and obtaining the erroneous words output from the first large-scale language model. The process involves obtaining one or more phoneme sequences of the pronunciation of the aforementioned incorrect word, and outputting word correction candidates for which the normalized phoneme distance between two phonemes in the aforementioned phoneme sequence is less than or equal to a predetermined threshold. The process involves inputting the aforementioned erroneous word, the word correction candidate output for the erroneous word, and a prompt instructing the user to select the word correction candidate to replace the erroneous word into a second large-scale language model, and then reflecting the word correction candidate output from the second large-scale language model into the speech recognition text and outputting it. An information processing program that executes [something].

Images