JP2021032909A - Prediction device, prediction method and prediction program - Google Patents

Abstract

To increase the accuracy of predicting word intelligibility which is a quality evaluation scale of voice signals.SOLUTION: A word intelligibility prediction device 10 has: a voice recognition unit 11 which performs voice recognition for voice signals of a prediction object, by using a sound model 121 which outputs likelihood showing which phoneme corresponds to each frame of the inputted voice signals; and a word intelligibility prediction unit 16 which predicts word intelligibility which is a quality evaluation scale of the voice signals, on the basis of the voice recognition result by the voice recognition unit 11.SELECTED DRAWING: Figure 1

Description

The present invention relates to a prediction device, a prediction method and a prediction program.

Voice signal quality evaluation scales include word intelligibility and syllable intelligibility. Word intelligibility is an index value that indicates the percentage of meaningful words spoken and transmitted that were correctly heard, and is defined as the percentage of the number of words that the listener was able to hear correctly out of the number of words that the listener listened to. To. Syllable intelligibility is an index value that indicates the percentage of meaningless syllables uttered and transmitted that are correctly heard, and is defined as the percentage of the number of syllables that the listener was able to hear correctly. To.

As an evaluation of word intelligibility, SRT (Speech Reception Threshold) calculated from the recognition rate when a subject recognizes a word of a voice signal, listening effort obtained from a questionnaire on ease of recognition, and the like are known. However, subject experiments are costly, both economically and temporally. Therefore, a method of objectively measuring word intelligibility from a voice signal has been proposed.

As a method for objectively measuring word intelligibility, for example, speech intelligibility index (AI: Articulation Index), speech intelligibility index (SII: Speech Intelligibility Index), speech transmission index (STI: Speech Transmission Index), PESQ ( Calculation methods such as Perceptual Evaluation of Speech Quality) are used. However, since these calculation methods are calculations assuming a linear system, there is a problem that appropriate evaluation cannot be performed for signal conversion including non-linear signal processing.

Therefore, the short time objective intelligibility (STOI), the hearing-aid speech perception index (HASPI), etc. are used as voice signals so that they can be applied to some non-linear signal processing. It is often used as a quality evaluation measure. Furthermore, a Gammachirp Envelope Distortion Index (GEDI) that considers human auditory characteristics has also been proposed.

On the other hand, the performance of the automatic speech recognizer using deep learning is close to the performance of human hearing, and it is expected that the recognition rate can approximate the recognition rate obtained in the subject experiment. Therefore, instead of the subject experiment, a method of predicting the voice signal quality by using the recognition by the automatic voice recognition recognizer has been proposed.

As this method, a matrix test is performed with an automatic voice recognizer, in which a voice signal that reads aloud a sentence is presented and a word corresponding to a part of the voice signal is selected from correct text candidates, and the word intelligibility is determined from the correct answer rate. There is a method of predicting SRT, which is one of the above (see Non-Patent Document 1).

Constantin Spille, Stephan D. Ewert, Birger Kollmeier and Bernd T. Meyer, “Predicting speech intelligibility with deep neural networks”, Computer Speech & Language, Vol. 48, pp. 51-66, 2018.

In an automatic speech recognizer, it is common to improve the recognition rate by using as much as possible what can be used, such as prior knowledge of a language such as using a word dictionary.

On the other hand, since the voice signal quality is a characteristic of the voice signal itself, it is desirable to avoid factors such as language knowledge from affecting the recognition rate. As an influence of linguistic knowledge, for example, it is conceivable that the context before and after is a hint in word recognition, and that the recognition rate changes greatly depending on whether or not it is registered in the word dictionary.

Therefore, in predicting the quality of the presented voice signal by the automatic voice recognizer, there is a problem that not only the voice signal but also the word knowledge used affects the prediction of the word intelligibility. For example, the more intimate a word the listener is familiar with, the higher the intelligibility of the word and the easier it is to predict. In order to avoid this influence, the technique described in Non-Patent Document 1 is devised such as using a matrix test in which any correct answer candidate can be a correct answer with the same degree of plausibility without depending on the context. In other words, it is necessary to devise the design of the evaluation experiment so that the influence of intimacy does not affect the quality prediction.

As described above, in the voice signal quality prediction technology using the automatic voice recognizer described in Non-Patent Document 1, freely spoken voice, read-aloud voice of sentences that are not considered regarding the prior language information of the automatic voice recognizer, etc. Then, since the intimacy of words is not unified, there is a problem that it is difficult to obtain an accurate predicted value.

The present invention has been made in view of the above, and can improve the prediction accuracy of word intelligibility, which is a quality evaluation scale for voice signals, without requiring prior measures such as unifying the intimacy of words. It is an object of the present invention to provide a prediction device, a prediction method and a prediction program capable of providing a prediction device and a prediction program.

In order to solve the above-mentioned problems and achieve the object, the prediction device according to the present invention uses an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to. It is characterized by having a voice recognition unit that performs voice recognition for the voice signal to be predicted and a prediction unit that predicts the word intelligibility, which is a quality evaluation scale of the voice signal, based on the voice recognition result by the voice recognition unit. To do.

Further, the prediction method according to the present invention is a prediction method executed by a prediction device, and predicts using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to. It is characterized by including a step of performing voice recognition for a target voice signal and a step of predicting word intelligibility, which is a quality evaluation scale of the voice signal, based on the voice recognition result.

Further, the prediction program according to the present invention includes a step of performing voice recognition for the voice signal to be predicted by using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to. , A computer is made to perform a step of predicting word comprehension, which is a quality evaluation scale of a voice signal, based on a voice recognition result.

According to the present invention, it is possible to improve the prediction accuracy of word intelligibility, which is a quality evaluation scale for voice signals.

FIG. 1 is a diagram showing an outline of a configuration of a word intelligibility prediction device according to an embodiment. FIG. 2 is a diagram illustrating learning of the acoustic model and the phoneme language model shown in FIG. FIG. 3 is a diagram illustrating parameter adjustment of the prediction function of the word intelligibility prediction unit shown in FIG. FIG. 4 is a diagram illustrating processing of the word intelligibility prediction device shown in FIG. FIG. 5 is a flowchart showing a processing procedure of the word intelligibility prediction processing according to the embodiment. FIG. 6 is a diagram illustrating an evaluation experiment of the word intelligibility prediction device shown in FIG. FIG. 7 is a diagram showing an example of a computer in which a word intelligibility prediction device is realized by executing a program.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. The present invention is not limited to this embodiment. Further, in the description of the drawings, the same parts are indicated by the same reference numerals.

[Embodiment]
Embodiments of the present invention will be described. The present embodiment relates to a word intelligibility predictor that predicts word intelligibility obtained in a subject experiment based on the phoneme recognition rate of a speech recognizer.

First, the configuration of the word intelligibility prediction device according to the embodiment will be described. FIG. 1 is a diagram showing an outline of a configuration of a word intelligibility prediction device according to an embodiment. The word intelligibility prediction device 10 according to the embodiment predicts the word intelligibility based on the voice recognition rate for the input voice signal.

In the word comprehension prediction device 10, for example, a predetermined program is read into a computer or the like including a ROM (Read Only Memory), a RAM (Random Access Memory), a CPU (Central Processing Unit), etc., and the CPU executes the predetermined program. It is realized by executing. Further, the word intelligibility prediction device 10 has a communication interface for transmitting and receiving various information to and from other devices connected via a network or the like. For example, the word intelligibility prediction device 10 has a NIC (Network Interface Card) or the like, and communicates with other devices via a telecommunication line such as a LAN (Local Area Network) or the Internet. The word intelligibility prediction device 10 includes a voice recognition unit 11 and a word intelligibility prediction unit 16 (prediction unit).

The voice recognition unit 11 is an automatic voice recognizer that performs voice recognition for a voice signal to be predicted by using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to. .. The voice recognition unit 11 includes a phoneme output unit 12, a phoneme arrangement output unit 13, a phoneme recognition unit 14 (recognition unit), and a recognition rate calculation unit 15 (calculation unit).

The phoneme output unit 12 uses the acoustic model 121 to output phoneme candidates corresponding to each frame of the voice signal to be predicted.

The acoustic model 121 is a model that outputs which phoneme each frame of the input audio signal is likely to correspond to. The acoustic model 121 is a deep learning model. A deep learning model consists of an input layer into which signals enter, one or more intermediate layers that variously convert signals from the input layers, and an output layer that converts the signals of the intermediate layers into outputs such as probabilities. When a voice signal is input to the input layer, the acoustic model 121 outputs the probability of each phoneme indicating which phoneme each frame of the input voice signal is likely to correspond to from the output layer. To.

The phoneme arrangement output unit 13 uses the phoneme language model 131 to output phoneme arrangement candidates corresponding to the phoneme candidates output by the phoneme output unit 12.

The phoneme language model 131 is a model that outputs the plausibility of the arrangement of phonemes with respect to the input phoneme candidates. A phoneme language model such as a phoneme n-gram is applied to the phoneme language model 131 by calculating and learning the appearance frequency of a sequence of phonemes from the correct text.

The phoneme recognition unit 14 recognizes a phoneme sequence corresponding to a phoneme signal to be predicted based on a phoneme candidate output by the phoneme output unit 12 and a phoneme sequence candidate output by the phoneme arrangement output unit 13. The phoneme recognition unit 14 outputs a phoneme sequence (hereinafter, regarded as a word) from phoneme candidates and phoneme sequence candidates.

The recognition rate calculation unit 15 calculates the correct answer rate of the phoneme sequence recognized by the phoneme recognition unit 14. The recognition rate calculation unit 15 converts the correct text into words. The correct text is the original sentence in the case of the read-aloud voice of the sentence, and if the original voice is sufficiently clean, it may be manually transcribed. After that, the recognition rate calculation unit 15 collates the output phoneme series with the phoneme series of the correct answer text, and outputs the phoneme recognition correct answer rate. Recognition rate calculating unit 15, using equation (1), calculates the phoneme recognition accuracy rate _{P ACC.} In the equation (1), C is a correct phonetic prime number, S is a substitution phonetic prime number, I is an inserted phonetic prime number, and D is a deleted phonetic prime number.

The word intelligibility prediction unit 16 predicts the word intelligibility, which is a quality evaluation scale for voice signals, based on the voice recognition result by the voice recognition unit 11, and outputs the predicted value. The word intelligibility prediction unit 16 converts the phoneme recognition correct answer rate of the phoneme series calculated by the recognition rate calculation unit 15 into a predicted value of the word intelligibility by using a predetermined prediction function.

FIG. 2 is a diagram illustrating learning of the acoustic model 121 and the phoneme language model 131 shown in FIG. The parameters of the acoustic model 121 and the phoneme language model 131 are adjusted by learning a dataset of speech data and correct text.

As shown in FIG. 2, first, a clean audio signal data set Ds1 and a data set of the correct text are prepared. Then, the audio signal data is processed, a new audio signal is created, and the processed audio signal data set Ds2 is prepared by performing processing for adding various noises, speech enhancement processing, and the like to the clean audio signal.

The acoustic model 121 is made to learn the clean audio signal data set Ds1 and the processed audio signal data set Ds2 (step S2), and the parameters of the acoustic model 121 are adjusted. A conventional method is used for learning the acoustic model 121. For the specific procedure of the conventional method, refer to, for example, Tatsuya Kawahara, "Voice Recognition System Revised 2nd Edition", Ohmsha, 2016.

For the phoneme language model 131, the appearance frequency of the phoneme sequence is calculated from the correct text, the phoneme language model such as the phoneme Ngram is learned (step S1), and the parameters of the phoneme language model 131 are adjusted.

FIG. 3 is a diagram illustrating parameter adjustment of the prediction function of the word intelligibility prediction unit 16 shown in FIG. FIG. 4 is a diagram illustrating the processing of the word intelligibility prediction device 10 shown in FIG.

First, as a preparatory step, the word intelligibility prediction unit 16 is calibrated. Prepare the reference audio signal and its correct text for calibration. Then, a subject experiment is conducted using the illuminating voice signal, and the intelligibility such as the word recognition rate is evaluated. On the other hand, a reference audio signal is input to the word intelligibility prediction device 10 to output a phoneme recognition correct answer rate.

Subsequently, the parameters of the prediction function of the word intelligibility prediction unit 16 are adjusted based on the result of the subject experiment and the phoneme recognition correct answer rate by the word intelligibility prediction device 10 (step S3 in FIG. 3), and the predicted value is obtained. Match the results of the subject experiment. After adjusting the parameters of the prediction function, as shown in FIG. 4, as an actual prediction process, the audio signal to be predicted and the text thereof are input to the word intelligibility prediction device 10 to obtain an output of the word intelligibility prediction value. ..

[Prediction processing]
Next, the prediction process executed by the word intelligibility prediction device 10 will be described. FIG. 5 is a flowchart showing a processing procedure of the word intelligibility prediction processing according to the embodiment.

When the voice signal to be predicted is input, as shown in FIG. 5, the voice recognition unit 11 first determines whether or not there is unprocessed data (step S11). When there is unprocessed data (step S11: Yes), the voice recognition unit 11 reads the voice signal to be predicted (step S12) and performs voice recognition.

Specifically, the phoneme output unit 12 uses the acoustic model 121 to output phoneme candidates corresponding to each frame of the voice signal to be predicted (step S13). Subsequently, the phoneme arrangement output unit 13 uses the phoneme language model 131 to output the phoneme arrangement candidates corresponding to the phoneme candidates output by the phoneme output unit 12 (step S14). The phoneme recognition unit 14 recognizes a word corresponding to the voice signal to be predicted based on the phoneme candidate and the phoneme arrangement candidate (step S15), and the voice recognition unit 11 proceeds to step S11.

On the other hand, when there is no unprocessed data (step S11: No), the voice recognition unit 11 reads the correct answer text (step S16). Then, the recognition rate calculation unit 15 converts the correct answer text into words, compares all the words recognized by the phoneme recognition unit 14 with the words of the correct answer text, and calculates the phoneme recognition correct answer rate (step S17).

The word intelligibility prediction unit 16 converts the phonetic recognition correct answer rate of the word calculated by the recognition rate calculation unit 15 into the predicted value of the word intelligibility by using the prediction function to obtain the predicted value of the word intelligibility. Calculate (step S18). The word intelligibility prediction unit 16 outputs the predicted value of the word intelligibility (step S19), and ends the process.

[Evaluation experiment]
FIG. 6 is a diagram illustrating an evaluation experiment of the word intelligibility prediction device 10 shown in FIG. In the evaluation experiment, CSJ (The corpus of spontaneous Japanese) (for details, Sadaoki Furui, Kikuo Maekawa, and Hitoshi Isahara, “A japanese national project on sponta-neous speech corpus and processing technology) was used as the voice signal data set (training data). ”, In ASR2000-Automatic Speech Recognition: Challenges for the new Millenium ISCA Tutorial and Research Workshop (ITRW), pp. 244-248, 2000, and Kikuo Maekawa,“ CORPUS OF SPONTANEOUS JAPANESE: ITS DESIGN AND EVALUATION ”, In ISCA & IEEE Workshop on Spontaneous Speech Processing and Recognition, 2003). Here, the phoneme language model 131 was learned using the phoneme biggram obtained from the CSJ corpus.

In the evaluation experiment, a signal in which some intensity pink noise is added to this audio signal and a signal in which the audio signal to which pink noise is added are speech-enhanced are created as training data. Here, as speech enhancement, SS (spectral subtraction) (for details, see Michael Berouti, Richard Schwartz, and John Makhoul, “Enhancement of speech corrupted by acoustic noise”, In ICASSP'79. IEEE International Conference on Acoustics, Speech, and Signal Processing, Vol. 4, pp. 208-211. IEEE, 1979) and WF (Wiener filter) (for details, Masakiyo Fujimoto, Shinji Watanabe, and Tomohiro Nakatani, “Noise suppression with unsupervised joint speaker adaptation and noise) "mixture model estimation", In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4713-4716. See IEEE, 2012).

The acoustic model 121 was trained using a mixture of a clean audio signal, an audio signal with pink noise added, and a speech-enhanced audio signal.

As a data set (evaluation data) for evaluating word intelligibility, the familiarity-controlled word lists 2007 (FW07) (for details, see Shuichi Sakamoto, Naoki Iwaoka, Yoiti Suzuki, Shigeaki Amano, and Tadahisa Kondo, “Complementary relationship between familiarity and” SNR in word intelligibility test ”, Acoustical science and technology, Vol. 25, No. 4, pp. 290-292, 2004, and T Kondo, S Amano, S Sakamoto, and Y Suzuki,“ Family-controlled word lists 2007 (fw07) ”, The Speech Resources Consortium, National Institute of Informatics, Japan, 2007).

This dataset is divided by word intimacy, and only the one with the lowest intimacy is used in order to suppress the influence of word knowledge on the recognition rate. Similar to CSJ, FW07 is also subjected to pink noise addition and speech enhancement processing.

In this evaluation experiment, the word recognition rate from the subject experiment is used to calculate the word intelligibility. Then, the word intelligibility of the speech-enhanced voice signal is predicted by the word intelligibility prediction unit 16. The word intelligibility prediction unit 16 uses the linear function shown in the equation (2) as the conversion from the phoneme recognition correct answer rate of the speech recognition unit 11 to the word intelligibility.

Here, P _ASR is the phoneme recognition correct answer rate of the voice recognition unit 11, and SI _sub is the predicted value of the word intelligibility. The coefficients a and b of the linear function are set by using the least squares method from the phoneme recognition correct answer rate of the speech recognition unit 11 and the word intelligibility of the subject experiment of the speech signal to which pink noise is added. When a set of phoneme recognition correct answer rate and word intelligibility of the voice recognition unit 11 ( _PASR (i), SI _sub (i)), i = 1, 2, ..., N is given, the coefficient a, The value of b is estimated as in the following equations (3) and (4).

When the coefficients a and b were estimated using the data with pink noise of 3 dB, 0 dB, -3 dB, and -6 dB added, the equations (5) and (6) were obtained.

As speech enhancement, the predicted value of the word intelligibility predicted by the word intelligibility predictor 10 for the voice signal processed by SS and WF (predicted value of the objective word intelligibility) and the result of the subject experiment (subjective word intelligibility). ) And the mean squared error are shown in Table 1. ASR is the result of the word intelligibility predictor 10. For details on the calculation of GEDI, STOI, and HASPI, which are the conventional methods, see Katsuhiko Yamamoto, Toshio Irino, Shoko Araki, Keisuke Kinoshita, and Tomohiro Nakatani, "GEDI: Gammachirp Envelope Distortion Index for Predicting Intelligibility of Enhanced Speech", arXiv preprint arXiv: It is described in 1904.02096, 2019.

As shown in Table 1, the mean squared prediction error between the predicted value of the objective word intelligibility and the subjective word intelligibility was the smallest in ASR. That is, ASR had the highest prediction performance as compared with the conventional GEDI, STOI, and HASPI.

[Effect of Embodiment]
In the present embodiment, voice recognition is performed on the voice signal to be predicted by using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to, and the voice recognition result is used as a basis. In addition, the word comprehension level, which is a speech signal quality evaluation scale, is predicted. As shown in the evaluation experiment described above, according to the present embodiment, it is possible to improve the prediction accuracy of the word intelligibility as compared with the conventional STOI, HASPI and the recently proposed GEDI.

Here, the conventional automatic speech recognition device uses a word dictionary, a language model, or the like, and is easily affected by the context and prior knowledge of words in recognition. In order to eliminate such effects, it is necessary to use words that do not depend on the context of the context in the evaluation test, or to unify the intimacy of the words included in the utterances used in the test. Unless such pre-adjustment is made, the word intelligibility cannot be predicted accurately, and the accuracy of predicting the quality of the audio signal itself is lowered.

On the other hand, in the present embodiment, the speech recognition unit 11 uses a phoneme-level phoneme language model 131 called a phoneme Ngram, instead of information on the context before and after or language information such as a word dictionary. As a result, the voice recognition unit 11 can perform voice recognition without being affected by the context and prior knowledge of words, and the word intelligibility prediction unit 16 is also independent of linguistic information and can be used for various texts. It has become possible to predict the quality of voice signals.

That is, according to the present embodiment, it is possible to predict the word intelligibility that does not depend on the utterance content of the voice signal. In other words, according to the present embodiment, it is possible to predict the word intelligibility that does not depend on the intimacy of the word. For this reason, even if a word list with unified word intimacy is prepared in advance for the test, it is possible to approximate the result of the subject experiment more accurately than the conventional objective speech quality index. it can.

In the present embodiment, the case of predicting the word intelligibility as an objective evaluation index of voice quality has been described as an example, but the present invention is not limited to this. When syllable intelligibility is used as an objective evaluation index of voice quality, and when word recognition rate or character recognition rate is used as the recognition rate of the voice recognizer, the voice recognition result by the voice recognition unit 11 is obtained as in the present embodiment. It is possible to calculate the predicted value based on this. Specifically, a configuration in which the word intelligibility in the present embodiment is replaced with syllable intelligibility may be adopted. Alternatively, a configuration in which the phoneme recognition correct answer rate in the present embodiment is replaced with a character recognition correct answer rate or a word recognition correct answer rate may be adopted. In the character recognition correct answer rate, C in the above equation (1) is the number of correct characters, S is the number of replacement characters, I is the number of inserted characters, and D is the number of deleted characters. In the word recognition correct answer rate, C in the above equation (1) is the number of correct words, S is the number of replacement words, I is the number of inserted words, and D is the number of deleted words. Further, the word intelligibility in the present embodiment is defined as syllable intelligibility and the phoneme recognition correct answer rate is replaced with the character recognition correct answer rate, and the word intelligibility in this embodiment is defined as syllable intelligibility and the phoneme recognition correct answer rate is defined. It may be configured by replacing it with the word recognition intelligibility.

[System configuration, etc.]
Each component of each of the illustrated devices is a functional concept and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of the device is functionally or physically distributed in arbitrary units according to various loads and usage conditions. Can be integrated and configured. Further, each processing function performed by each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

Further, among the processes described in the present embodiment, all or part of the processes described as being automatically performed can be manually performed, or the processes described as being manually performed. It is also possible to automatically perform all or part of the above by a known method. In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[program]
FIG. 7 is a diagram showing an example of a computer in which the word intelligibility prediction device 10 is realized by executing a program. The computer 1000 has, for example, a memory 1010 and a CPU 1020. The computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these parts is connected by a bus 1080.

The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, the display 1130.

The hard disk drive 1090 stores, for example, an OS (Operating System) 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each process of the word intelligibility prediction device 10 is implemented as a program module 1093 in which a code that can be executed by a computer is described. The program module 1093 is stored in, for example, the hard disk drive 1090. For example, a program module 1093 for executing a process similar to the functional configuration in the word intelligibility prediction device 10 is stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

Further, the setting data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 as needed, and executes the program.

The program module 1093 and the program data 1094 are not limited to those stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). Then, the program module 1093 and the program data 1094 may be read by the CPU 1020 from another computer via the network interface 1070.

Although the embodiment to which the invention made by the present inventor is applied has been described above, the present invention is not limited by the description and the drawings which form a part of the disclosure of the present invention according to the present embodiment. That is, all other embodiments, examples, operational techniques, and the like made by those skilled in the art based on the present embodiment are included in the scope of the present invention.

10 Word intelligibility prediction device 11 Speech recognition unit 12 Phoneme output unit 13 Phoneme alignment output unit 14 Phoneme recognition unit 15 Recognition rate calculation unit 16 Word intelligibility prediction unit 121 Acoustic model 131 Phoneme alignment language model

Claims (6)

A voice recognition unit that performs voice recognition for the voice signal to be predicted by using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to.
Based on the voice recognition result by the voice recognition unit, a prediction unit that predicts word intelligibility, which is a quality evaluation scale for voice signals, and a prediction unit.
A prediction device characterized by having. The voice recognition unit
Using the acoustic model, a phoneme output unit that outputs phoneme candidates corresponding to each frame of the voice signal to be predicted, and a phoneme output unit.
Using a phoneme language model that outputs the plausibility of phoneme arrangement for the input phoneme candidates, the phoneme arrangement output unit that outputs the phoneme arrangement candidates corresponding to the phoneme candidates output by the phoneme output unit. When,
A recognition unit that recognizes a phoneme sequence corresponding to the voice signal to be predicted based on the phoneme candidate and the phoneme sequence candidate, and a recognition unit.
The prediction device according to claim 1, wherein the prediction device has. The voice recognition unit
A calculation unit that calculates the phoneme recognition accuracy rate of the phoneme series recognized by the recognition unit, and
Have,
The second aspect of claim 2, wherein the prediction unit converts the phoneme recognition correct answer rate of the phoneme series calculated by the calculation unit into a predicted value of the word intelligibility by using a predetermined prediction function. Predictor. The prediction device according to any one of claims 1 to 3, wherein the acoustic model is a deep learning model. It is a prediction method executed by the prediction device.
The process of performing voice recognition for the voice signal to be predicted using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to.
The process of predicting word intelligibility, which is a quality evaluation scale for voice signals, based on the voice recognition results,
A prediction method characterized by including. Using an acoustic model that outputs which phoneme each frame of the input voice signal is likely to correspond to, the step of performing voice recognition for the voice signal to be predicted, and
Steps to predict word intelligibility, which is a quality evaluation scale for voice signals, based on voice recognition results,
A prediction program that lets your computer run.

Images