JP5544575B2 - Spoken language evaluation apparatus, method, and program - Google Patents

Description

  The present invention relates to a spoken language evaluation apparatus, method, and program, and more particularly, to a spoken language evaluation apparatus, method, and program for evaluating the type of language indicated by an input speech signal.

  Conventionally, the type of language indicated by an audio signal is identified from the audio signal, and many techniques have been proposed for that purpose (for example, Patent Literature 1, Patent Literature 2, Non-Patent Literature 1, and Non-Patent Literature 1). (See Patent Document 2). As a technology for identifying the type of language indicated by such an audio signal, there are mainly a technique that utilizes not only sound information but also text level grammar, and a technique that utilizes only phonetic information and features of phoneme level. Can be classified.

  As a technique using text-level grammar, for example, in the technique described in Patent Document 1, language recognition and analysis are performed by natural language analysis processing using a vocabulary grammar model, semantic rules, and the like. In addition, as methods utilizing the features of phoneme levels, many methods for classifying languages in consideration of similarities to sounds included in each language such as vowels have been proposed. For example, in the technique described in Patent Document 2, language recognition is performed using phoneme information instead of text, assuming several phonetic alphabets as prior knowledge.

JP-A-8-106374 JP 2001-109490 A

Zissman, M.A. "Comparison of four approaches to automatic language identification of telephone speech," IEEE Trans. On Speech and Audio Processing, Vol.4, No.1, pp. 31-44, Jan. 1996. Yeshwant K. et.al "Reviewing Automatic Language Identification," IEEE Signal Processing Magazine, pp. 33-41, Oct. 1994

  However, the technique using the text level grammar has a problem that it is difficult to apply to a language which has no characters and the grammar is not analyzed. For example, the technique described in Patent Document 1 requires a natural language analysis process and cannot be applied to a language in which no character language exists.

  Further, when utilizing the features of phoneme level, for example, as in the technique described in Patent Document 2, it is difficult to apply to a language that requires prior knowledge and does not have many characters that have not been analyzed. There is a problem that.

  The present invention has been made to solve the above-described problems, and evaluates the type of language indicated by an input speech signal without performing conversion to a text-level language expression and without requiring prior knowledge. An object of the present invention is to provide a spoken language evaluation apparatus, method, and program that can be used.

  In order to achieve the above object, the spoken language evaluation apparatus of the present invention includes an extraction means for extracting evaluation feature information from an evaluation speech signal whose language type is unknown, and a plurality of learning speeches whose language types are known. The phoneme representation for each language type represented by the basis vector for each phoneme obtained by non-negative matrix decomposition of the learning feature information for each language type extracted from the signal, and extracted by the extraction means Based on the evaluation feature information, a blending ratio calculating means for calculating a blending ratio indicating a base vector ratio of each phoneme blended in the evaluation feature information for each language type, the evaluation feature information, , Corresponding to the evaluation feature information based on the similarity between each of the information indicated by the product of the blending ratio for each language type calculated by the blending ratio calculating means and the phoneme representation for each language type. Audio signal for evaluation It is configured to include an evaluation means for evaluating the type of to language, the.

  According to the speech language evaluation apparatus of the present invention, the extraction unit extracts the evaluation feature information from the evaluation speech signal whose language type is unknown. Further, by performing non-negative matrix decomposition on learning feature information for each language type extracted from a plurality of learning speech signals whose language types are known, each language type represented by a base vector for each phoneme is obtained. Phoneme expression is obtained in advance. Then, based on the phoneme representation for each type of language obtained in advance and the evaluation feature information extracted by the extraction unit, the blending ratio calculation unit calculates the basis vector of each phoneme blended in the evaluation feature information. A blending ratio indicating a ratio is calculated for each language type. Then, the evaluation means is based on the similarity between the evaluation feature information and each information indicated by the product of the blending ratio for each language type calculated by the blending ratio calculating means and the phoneme representation for each language type. The language type indicated by the evaluation audio signal corresponding to the evaluation feature information is evaluated.

  In this way, it is represented by the product of the phoneme representation for each language type obtained by non-negative matrix decomposition of the learning speech signal and the blending ratio calculated based on the phoneme representation and the evaluation feature information. Because the type of language indicated by the evaluation speech signal is evaluated based on the similarity between the information and the evaluation feature information, it is input without conversion to a text-level language expression and without prior knowledge. The type of language indicated by the audio signal can be evaluated.

  The phoneme expression may be a phoneme expression having a time-series structure. Accordingly, it is possible to evaluate the type of language indicated by the input voice signal in consideration of a subtle change in phonemes in a continuous change in sound.

  Further, the evaluation means identifies the language type corresponding to the phoneme expression when the similarity is the highest as the language type indicated by the evaluation speech signal, or for each language type Based on the similarity, a language phylogenetic tree showing the systematic relationship between the types of languages can be created.

  In addition, the blending ratio calculation means is provided for at least one of the language type and the sex and age obtained from the learning feature information extracted from the learning speech signal in which at least one of the sex and age of the speaker is known. Based on the phoneme expression, the blending ratio can be calculated for at least one of language type, sex, and age.

  Further, the extraction means extracts learning feature information for each language type from the plurality of learning speech signals, and non-negative matrix decomposition of the learning feature information for each language type extracted by the extraction means Thus, a phoneme expression calculating means for calculating a phoneme expression for each language type can be included.

  The spoken language evaluation method of the present invention is a spoken language evaluation method in a spoken language evaluation apparatus including an extraction unit, a blending ratio calculation unit, and an evaluation unit, and the extraction unit has an unknown language type. The feature information for evaluation is extracted from the speech signal for evaluation, and the blending ratio calculation means performs non-negative matrix decomposition of the learning feature information for each language type extracted from a plurality of learning speech signals whose language types are known. Each of the blended features in the evaluation feature information based on the phoneme representation for each language type represented by the basis vector for each phoneme obtained by the above and the evaluation feature information extracted by the extraction means A blending ratio indicating a ratio of basis vectors of phonemes is calculated for each language type, and the evaluation unit includes the evaluation feature information, a blending ratio for each language type calculated by the blending ratio calculating unit, and the Language seed Based on the similarity between the information each represented by the product of the phoneme transcription for each, it is a method of evaluating the type of language indicated by the audio signal corresponding to the evaluation feature information.

  Further, in the spoken language evaluation method in the spoken language evaluation apparatus further including a phoneme expression calculating unit, the extracting unit extracts learning feature information for each language type from the plurality of learning speech signals, and the phoneme expression The calculation means calculates a phoneme expression for each language type by performing non-negative matrix decomposition on the learning feature information for each language type extracted by the extraction means.

  The spoken language evaluation program of the present invention is a program for causing a computer to function as each means constituting the above-described speech language evaluation apparatus.

  As described above, according to the speech language evaluation apparatus, method, and program of the present invention, the phoneme expression for each language type obtained by non-negative matrix decomposition of the learning speech signal, and the phoneme expression and evaluation thereof In order to evaluate the type of language indicated by the evaluation speech signal based on the similarity between the information indicated by the product of the blending ratio calculated based on the feature information and the feature information for evaluation, the text level language expression is used. Thus, there is an effect that it is possible to evaluate the type of language indicated by the input voice signal without performing the above conversion and without requiring prior knowledge.

It is the schematic which shows the structure of the spoken language evaluation apparatus which concerns on 1st Embodiment. It is an image figure of nonnegative matrix decomposition | disassembly. It is a graph which shows an example of the phoneme expression of Chinese. It is a graph which shows an example of the phoneme expression of Spanish. It is a flowchart which shows the content of the learning process routine in the spoken language evaluation apparatus which concerns on 1st Embodiment. It is a flowchart which shows the content of the evaluation process routine in the spoken language evaluation apparatus which concerns on 1st Embodiment. It is an image figure of nonnegative matrix decomposition | disassembly with respect to a time-sequential phoneme expression. It is a graph which shows an example of the phoneme expression of the time series of English. It is a graph which shows an example of the phoneme expression of the time series of German. It is a graph which shows an example of the phoneme expression of the time series of Swedish. It is a graph which shows an example of the phoneme expression of the time series of French. It is a graph which shows an example of the similar value with respect to a certain audio | voice signal. It is a figure which shows an example of the output of a language phylogenetic tree. It is a graph which shows an example of the language classification using the time-sequential phoneme expression.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  A spoken language evaluation apparatus 1 according to the first embodiment includes a CPU, a RAM, and a ROM that stores a program for executing a spoken language evaluation processing routine including a learning process and an evaluation process described later. It consists of

  As shown in FIG. 1, this computer functionally includes an audio signal input unit 11 for inputting an audio signal, a feature information extraction unit 12 for extracting feature information from the audio signal, and a language for pre-learning. Phoneme expression calculation unit 13 for calculating phoneme expression for feature information obtained for each type and sex, phoneme expression storage unit 14 for storing phoneme representation for each language type and sex, and for language evaluation A phoneme blending ratio calculation unit 15 that calculates a blending ratio for the feature information obtained from the feature information extraction unit 12 using each of the phoneme representations for each language type and gender stored in the phoneme representation storage unit 14; Analyzing the phoneme blending ratio for each language type and evaluating the similarity to each language, and the display control unit 17 for controlling the evaluation result to be displayed on the display device Can be represented by the included configuration .

  Also, the speech signal input unit 11, the feature information extraction unit 12, the phoneme expression calculation unit 13, and the phoneme representation storage unit 14 function as the learning unit 2, and the speech signal input unit 11, the feature information extraction unit 12, the phoneme blending ratio calculation. The unit 15, the language similarity evaluation unit 16, and the display control unit 17 function as the evaluation unit 3. That is, the audio signal input unit 11 and the feature information extraction unit 12 are used in common by the learning unit 2 and the evaluation unit 3.

  The audio signal input unit 11 receives a digitized audio signal from an input device such as an electronically recorded file or a microphone. In the learning stage, a speech signal (learning speech signal) having a known language type and gender (gender-specific) of the speaker is input. In the evaluation stage, a speech signal (evaluation speech signal) whose language type is unknown and whose speaker's gender is known or unknown is input.

  The feature information extraction unit 12 extracts feature information from the digitized audio signal obtained from the audio signal input unit 11. In the present embodiment, a case where a mel spectrum is extracted as feature information will be described. The feature information may be extracted depending on what is used for the phoneme expression and its identification method (for example, spectrum and principal component analysis (PCA), mel cepstrum and vector quantization, etc.). In the learning stage, a mel spectrum is extracted from the learning speech signal for each language type and sex, and this is used as learning feature information. Further, in the evaluation stage, a mel spectrum is extracted from the evaluation audio signal, and this is used as evaluation feature information.

The phoneme expression calculation unit 13 analyzes the mel spectrum for each language type and sex extracted as the learning feature information by the feature information extraction unit 12, and extracts phoneme structures that repeatedly appear in the speech signal. In such a method, for example, non-negative matrix factorization (NMF) suitable for handling non-negative information such as sound can be used (for example, “DD Lee, HS Seung, “Learning the part of objects by non-negative matrix factorization,” Nature Vol. 401, pp. 788-791, 1999. ”). NMF is applied to automatic transcription and separation of sound sources from monaural mixed signals (for example, “P. Smaragdis, JC Brown,“ Non-Negative Matrix Factorization for Music Transcription, ”ln Proc. 2003 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA2003), pp. 177-180,2003. ”And“ T. Virtanen, “Monaural sound source separation by nonnegative matrix factorization with temporal continuity and sparseness criteria,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, pp. 1066-1074, 2007 ”)
FIG. 2 shows an image in which an audio signal is decomposed into phonemes by NMF. In the figure, Y is the mel spectrum extracted by the feature information extraction unit 12, and H in the figure is a phoneme expression (arranged base vectors for each phoneme. The phoneme base vectors are also simply referred to as “phonemes” below. ), U in the figure represents a blending ratio indicating how much each phoneme is blended with Y. Appropriate phoneme representation H and blending ratio U can be obtained by minimizing the difference between Mel spectrum Y and the product of phoneme representation H and blending ratio U by repetitive calculation by NMF. In the evaluation stage, since only the phoneme expression H is used, the obtained phoneme expression H is output.

  Here, KL (Kullback-Leibler) -divergence is used as a distance measure used to minimize the difference between the mel spectrum Y and the product of the phoneme expression H and the blending ratio U in NMF. Instead of KL-divergence, the Itakura Saito distance or the Euclidean distance may be used. FIG. 3 shows a Chinese phoneme expression created using the mel spectrum as an input and KL-divergence as a distance scale. In the same figure, 10 phonemes are arranged in the horizontal direction, and the phoneme expression is expressed by using the vertical axis of each phoneme as the frequency and the horizontal axis as its strength. A similar phoneme expression in Spanish is shown in FIG. In Spanish and Chinese, there are similar phonemes and phonemes that are not. Such a feature is assumed to correspond to the difference in the type of vowel in each language.

  The phoneme expression storage unit 14 stores the phoneme expression H output from the phoneme expression calculation unit 13 for each language type for each gender. Here, the phoneme expression is stored for each language type for each gender, but depending on the feature information and the definition of the distance, men and women may be grouped or may be further classified by age.

  The phoneme blending ratio calculation unit 15 receives the mel spectrum extracted by the feature information extraction unit 12 as the evaluation feature information. In the NMF, the mel spectrum Y of the audio signal can be approximately expressed by the product of the phoneme expression H and the blending ratio U as shown in the upper equation of FIG. Based on this approximate expression, the phoneme blending ratio calculation unit 15 determines the language type based on the input mel spectrum Y and the phoneme expression H for each language type stored in the phoneme expression storage unit 14 for each gender. The blending ratio U is calculated every time.

  Here, when the gender of the speaker of the evaluation speech signal is known, the phoneme corresponding to the gender of the speaker out of the phoneme representation H for each type of language for each gender stored in the phoneme representation storage unit 14. Using the expression H, the blending ratio U is calculated. When the gender of the speaker is unknown, the blending ratio U is calculated for each language type and gender using all the stored phoneme expressions H.

  The language similarity evaluation unit 16 calculates the product of the blending ratio U calculated by the phoneme blending ratio calculation unit 15 and the phoneme representation H stored in the phoneme representation storage unit 14 for each language type, and extracts feature information. The similarity with the mel spectrum Y of the audio signal for evaluation output from the unit 12 is calculated. The similarity can be a difference (distance) between Y and the product of U and H. Thereby, the similarity between the language type indicated by the input speech signal and the phoneme representation of each language type is expressed as a distance. The language type corresponding to the phoneme expression H used for the calculation when the distance is the shortest is evaluated as the language type most similar to the language type indicated by the evaluation speech signal.

  Moreover, the language similarity evaluation part 16 may obtain | require the language phylogenetic tree with respect to the input audio | voice signal as an evaluation result for the systematization of a language using the calculated similarity. As a method of creating a language tree, a group average method (UPGMA: Unweighted Pair-Group Method using Average) or the like can be used. UPGMA is a method of creating a step-by-step language phylogenetic tree, and is a method of repeating a process of combining two languages having the minimum distance. In calculating the distance to the combined language group, the average value of the distance to each language in the group is used.

  If the gender of the speaker of the evaluation speech signal is unknown, both the similarity when using the male phoneme expression and the similarity when using the female phoneme expression are calculated. The gender corresponding to the phoneme expression having the higher degree is preferably obtained as an evaluation result.

  The display control unit 17 performs control so that the evaluation result by the language similarity evaluation unit 16 is displayed on the display device. For example, text indicates what language type is most similar to the language type indicated by the speech signal for evaluation, and displays the similarity between the speech signal for evaluation and the phoneme representation of each language type as a bar graph You can do it. In addition, when the language phylogenetic tree is obtained, the obtained language phylogenetic tree is preferably displayed.

  Here, the case where the evaluation result is displayed on the display device has been described. However, the evaluation result may be output by voice using a voice output device. For example, what is the most similar language type to the language type indicated by the evaluation audio signal can be displayed by voice, or the voice in the learning data of the most similar language type can be output.

  Next, the operation of the spoken language evaluation apparatus 1 according to the first embodiment will be described. Prior to the evaluation process for evaluating the type of language indicated by the evaluation speech signal, a learning process routine shown in FIG. 5 is executed.

  In step 100, a digitized learning audio signal is input from an electronically recorded file or input device such as a microphone.

  Next, in step 102, a mel spectrum is extracted as learning feature information from the learning speech signal input in step 100. The feature information for learning extracted here is feature information for each language type and sex.

  Next, in step 104, the mel spectrum extracted in step 102 is decomposed into phonemes by NMF for each language type and gender to obtain phoneme expression H and blending ratio U.

  Next, in step 106, the phoneme expression H calculated in step 104 is stored in the phoneme expression storage unit 14 for each language type for each gender, and the process is terminated.

  Then, the above-described learning process routine is executed, and the evaluation process routine shown in FIG. 6 is executed in a state where the phoneme expression storage unit 14 stores the phoneme expression H for each language type for each gender.

  In step 120, an audio signal for evaluation is input. Assume that the gender of the speaker is known from the evaluation audio signal input here.

  Next, in step 122, a mel spectrum is extracted as evaluation feature information from the evaluation speech signal input in step 120 by the same process as in step 102 of the learning process.

  Next, at step 124, the mel spectrum Y extracted from the speech signal for evaluation at step 122 and the phoneme representation H for each language type corresponding to the gender of the speaker stored in the phoneme representation storage unit 14. Based on this, the blending ratio U is calculated for each language type.

  Next, in step 126, the product of the blending ratio U for each language type calculated in step 124 and the phoneme representation H for each language type corresponding to the gender of the speaker stored in the phoneme representation storage unit 14. Are calculated, and the degree of similarity with the mel spectrum Y extracted in step 122 is calculated. The language type corresponding to the phoneme expression H used for the calculation when the similarity is the highest is evaluated as the language type most similar to the language type indicated by the evaluation speech signal. Further, using the calculated similarity, a systematic tree for the input evaluation speech signal is obtained as an evaluation result for systematization of language types.

  Next, in step 128, the evaluation result in step 126 is displayed on the display device, and the process is terminated.

  As described above, according to the speech language evaluation apparatus of the first embodiment, when the learning feature information extracted from the learning speech signal is expressed by phoneme expression and its blending ratio by non-negative matrix division For each language type, and based on the evaluation feature information extracted from the evaluation speech signal and the stored phoneme representation for each language type, the blending ratio is determined for each language type. Based on the similarity between the calculated feature information for evaluation and the product of the stored phoneme representation for each language type and the calculated blending ratio, the language type indicated by the audio signal for evaluation is the type of language Evaluate if it is similar to In this way, it is possible to evaluate the type of language indicated by the input speech signal using only the speech signal without performing conversion to a text-level language expression and without requiring prior knowledge.

  In addition, language phylogenetic trees can be obtained by using the similarity between the language type indicated by the speech signal for evaluation and the type of each language, and new cultural and historical knowledge about the relationship between language types is also expected. it can.

  Next, a second embodiment will be described. The spoken language evaluation apparatus according to the second embodiment is different from the first embodiment in that the phoneme expression calculation unit 13 uses time-series phoneme expression, and this point will be described.

  Language characteristics are classified according to the type of vowel, etc., but especially when the volume of each phoneme such as a vowel changes continuously, such as continuous sounds, it continuously changes from the previous phoneme to the subsequent phoneme. In the process, there are situations where it becomes difficult to recognize (for example, a change in sound from “good” to “good” in “good morning”). In the second embodiment, the kind of language is evaluated in consideration of such subtle changes in sound in continuous sounds.

  As shown in FIG. 7, the phoneme representation calculation unit 13 according to the second embodiment performs non-negative matrix decomposition (NMFD), for example, “Paris Smaragdis, “Non-negative Matrix Factor Deconvolution; Extraction of Multiple Sound Sources from Monophonic Inputs,” Independent Component Analysis and Blind Signal Separation, Lecture Notes in Computer Science, 2004, Volume 3195/2004, 494-499 ”) Compute phoneme representation of.

  8 to 11 show time series phoneme expressions of English, German, Swedish, and French calculated by NMFD. Here, the phoneme expression is such that 12 phonemes are arranged side by side. Within each phoneme, five time-series changes are represented horizontally, and the frequency represents frequency. The darker each square inside, the stronger the value. That is, each phoneme represents the time transition of the mel spectrum.

  Regarding the learning process and evaluation process in the second embodiment, the phoneme expression is calculated by NMF in the learning process and evaluation process in the first embodiment, and the time-series phoneme expression is calculated by using the NMFD. Since only the point of calculation is different, the description is omitted.

  As described above, according to the spoken language evaluation apparatus of the second embodiment, in addition to the effects of the first embodiment, the subtle phoneme change in the continuous change of sound is also taken into consideration. The type of language indicated by the evaluation audio signal can be evaluated.

  Here, in order to explain the effect of the present invention, an example of an evaluation result will be described.

  A phoneme expression is created from a speech corpus of 21 languages, and an evaluation result of comparing similarities for a Japanese input of a woman is shown in FIG. In order to convert the distance into a similar value, exp (−distance value) is set as the similar value. As shown in the figure, it was confirmed that the similarity value with Japanese was the highest, and the language type indicated by the input voice signal could be correctly identified as Japanese.

  FIG. 13 shows the result of repeating the same method between languages and creating a language phylogenetic tree by the UPGMA method. The evaluation results in the figure do not necessarily match the linguistic phylogenetic tree of linguistic classification, but reflect the geographical proximity, suggesting some relationship between phonemes and language types. It seems that

  Next, FIG. 14 shows the results of measuring distances for four languages of English, German, Swedish, and French using NMFD. The vertical axis represents learned phoneme changes, and the horizontal axis represents the language type of the input speech. The black in each square is more similar, and the whiter the difference is. As can be seen from the figure, the type of language that is most similar to all input language types is the correct language type, and it can be seen that identification using NMFD is effective. In addition, Swedish phonemes are very different from other languages, and German phonemes are similar to other languages. In general, Swedish and French have more types of vowels than English, and German has fewer types of vowels. It seems to be influenced by this trend.

  In the above-described embodiment, the case where the learning unit and the evaluation unit are configured by one computer has been described. However, the learning unit and the evaluation unit may be configured by separate computers.

  The present invention is not limited to the above-described embodiments, and various modifications and applications can be made without departing from the gist of the present invention.

  For example, although the above-described spoken language evaluation apparatus has a computer system therein, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. .

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 1 Spoken language evaluation apparatus 2 Learning part 3 Evaluation part 11 Speech signal input part 12 Feature information extraction part 13 Phoneme expression calculation part 14 Phoneme expression storage part 15 Phoneme combination ratio calculation part 16 Language similarity evaluation part 17 Display control part

Claims (8)

Extraction means for extracting evaluation feature information from an evaluation speech signal whose language type is unknown;
For each language type represented by a basis vector for each phoneme obtained by non-negative matrix decomposition of learning feature information for each language type extracted from a plurality of learning speech signals with known language types Based on the phoneme expression and the evaluation feature information extracted by the extraction means, a combination that calculates a combination ratio indicating the ratio of the basis vectors of each phoneme included in the evaluation feature information for each language type A ratio calculation means;
Based on the similarity between the feature information for evaluation and information indicated by the product of the blending ratio for each language type calculated by the blending ratio calculating means and the phoneme representation for each language type, the evaluation Evaluation means for evaluating the type of language indicated by the evaluation audio signal corresponding to the feature information;
Spoken language evaluation device including   The spoken language evaluation apparatus according to claim 1, wherein the phoneme expression is a phoneme expression having a time-series structure.   The evaluation means identifies the language type corresponding to the phoneme expression when the similarity is the highest as the language type indicated by the evaluation speech signal, or the similarity for each language type The spoken language evaluation apparatus according to claim 1 or 2, wherein a language phylogenetic tree indicating a systematic relationship between language types is created based on the language.   The blending ratio calculation means includes at least one of a speaker's gender and age, and a phoneme expression for each language type and gender and age obtained from learning feature information extracted from a learning speech signal. The spoken language evaluation apparatus according to claim 1, wherein the blending ratio is calculated for each of at least one of language type, sex, and age based on the language. The extraction means extracts learning feature information for each type of language from the plurality of learning audio signals,
The phoneme expression calculation means for calculating the phoneme expression for each language type by performing non-negative matrix decomposition on the learning feature information for each language type extracted by the extraction means. The spoken language evaluation apparatus according to claim 1. A spoken language evaluation method in a spoken language evaluation apparatus including an extraction unit, a blending ratio calculation unit, and an evaluation unit,
The extracting means extracts evaluation feature information from an evaluation speech signal whose language type is unknown,
The blending ratio calculating means is represented by a basis vector for each phoneme obtained by non-negative matrix decomposition of learning feature information for each language type extracted from a plurality of learning speech signals with known language types. Based on the phoneme representation for each type of language and the evaluation feature information extracted by the extraction means, a blending ratio indicating a ratio of basis vectors of each phoneme blended in the evaluation feature information is expressed in the language. For each type of
The evaluation means is based on the similarity between the evaluation feature information and each information indicated by the product of the mixture ratio for each language type calculated by the combination ratio calculation means and the phoneme expression for each language type. A speech language evaluation method for evaluating a language type indicated by an evaluation speech signal corresponding to the evaluation feature information. The spoken language evaluation apparatus further includes a phoneme expression calculation unit,
The extraction means extracts learning feature information for each type of language from the plurality of learning audio signals,
7. The phonetic language evaluation according to claim 6, wherein the phoneme expression calculation unit calculates a phoneme expression for each language type by performing non-negative matrix decomposition on learning feature information for each language type extracted by the extraction unit. Method.   The spoken language evaluation program for functioning a computer as each means which comprises the spoken language evaluation apparatus of any one of Claims 1-5.