Foreign language learning reading quality analysis system and method based on voice interaction

Abstract

The invention provides a foreign language learning loud-reading quality analysis system and method based on voice interaction, and the method comprises the steps: collecting a loud-reading audio signal of a learner, extracting a fundamental frequency contour and a spectrum envelope feature through a Mel-frequency cepstral coefficient algorithm, and obtaining a preliminary voice feature set; aligning duration information and energy distribution of the learner audio and the standard reference audio by adopting a dynamic time warping algorithm according to the preliminary voice feature set, and determining an aligned deviation vector; processing the combination of rhythm adjustment indexes and the preliminary voice feature set by adopting a long-short-term memory network to obtain a standardized feature vector representation form; aiming at the standardized feature vector representation form, through calculating the Euclidean distance between the standardized feature vector representation form and a standard pronunciation vector, determining a cross-language comparison difference degree; and generating an adaptive improved sequence according to the targeted feedback data to obtain a final reading quality analysis result. According to the method, through fusion of rhythm adjustment and cross-language difference analysis, the accuracy and personalized improvement effect of reading quality evaluation are remarkably improved, and efficient technical support is provided for language learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a foreign language learning reading quality analysis system and method based on voice interaction. BACKGROUND

[0002] Foreign language learning is an important way to improve personal ability and international communication, especially in the context of globalization, mastering fluent foreign language oral becomes a key skill. Reading as a core link of foreign language learning, not only helps learners master pronunciation, but also improves language sense and expression ability. However, the existing foreign language learning reading quality analysis system has significant limitations in technical implementation. Existing solutions often rely on a single audio processing method, which is difficult to fully capture the subtle features of learners' reading. For example, many systems only focus on the correctness of pronunciation, ignoring the prosodic and rhythmic changes of speech, resulting in inaccurate feedback results and failing to provide personalized improvement suggestions for learners. In addition, existing methods often lack flexibility and adaptability in dealing with the diversity of different languages and accents, making it difficult to meet the learning needs in multiple scenarios.

[0003] In this field, the core challenge is how to extract multi-level speech features from learners' reading audio and convert them into standardized data form to support high-quality analysis and evaluation. The first technical factor is the multi-level feature extraction of speech signal. Learners' reading audio contains fundamental frequency contour, spectral envelope, duration information and energy distribution, etc. These features together determine the naturalness and accuracy of reading. However, current technology often has difficulty in efficient integration when dealing with these features, resulting in incomplete analysis results. For example, a learner may pronounce accurately when reading an English sentence, but the intonation fluctuation does not conform to the habits of native speakers, and existing systems cannot effectively identify this prosodic deviation. Another related factor is the standardized representation of features. Different learners' speech features vary due to language background and pronunciation habits, and how to convert these features into a unified data representation form becomes a bottleneck for subsequent analysis. For example, in cross-language learning scenarios, the system needs to compare the differences between learners and standard pronunciation, but the lack of standardized feature vectors will lead to inaccurate comparative evaluation.

[0004] Therefore, how to extract multi-level speech features from learners' reading audio and construct standardized feature vector representation has become a key problem for precise reading quality analysis. SUMMARY

[0005] The present application provides a foreign language learning reading quality analysis system and method based on voice interaction, mainly including:

[0006] Collecting learners' reading audio, extracting fundamental frequency contour and spectral envelope features using Mel frequency cepstral coefficient algorithm to obtain a preliminary set of speech features;

[0007] The dynamic time warping algorithm is used to align the duration and energy distribution of learner audio with that of standard audio to determine the bias vector;

[0008] Extract intonation fluctuation and rhythmic sub-features from the deviation vector, and mark intonation fluctuation exceeding the threshold as rhythmic deviation to obtain rhythmic adjustment index;

[0009] By combining prosodic adjustment indices and preliminary speech feature sets, standardized feature vectors are obtained through long short-term memory network processing.

[0010] Calculate the Euclidean distance between this vector and the standard pronunciation vector to determine the degree of difference in cross-linguistic comparison;

[0011] If the difference exceeds the threshold, extract personalized sub-vectors from the standardized feature vectors, determine the corresponding rhythm and cadence adjustment path, and obtain targeted feedback data.

[0012] Based on the feedback data, an adaptive improvement sequence is generated, and the final reading quality analysis results are obtained.

[0013] As an embodiment of the present invention, the step of extracting a preliminary speech feature set includes:

[0014] The acquired audio was subjected to short-time Fourier transform to generate a time-frequency graph. The fundamental frequency profile and spectral envelope data were extracted using the Mel frequency cepstral coefficient algorithm to obtain the first feature dataset.

[0015] If the dimensionality of the dataset exceeds the threshold, principal component analysis is used to reduce the dimensionality to obtain the second feature dataset, which is then integrated to generate the final speech feature set.

[0016] As one embodiment of the present invention, the step of determining the alignment deviation vector includes:

[0017] Feature matrices containing duration and energy distribution are extracted from learners and standard audio, and time series alignment is performed using a dynamic time warping algorithm to obtain a time mapping path.

[0018] The first deviation vector is obtained based on the path calculation difference. It is determined that the matching accuracy is insufficient. After adjusting the feature matrix, the alignment is re-aligned to obtain the optimized deviation vector.

[0019] As one embodiment of the present invention, the steps for obtaining the prosody adjustment index include:

[0020] Extract intonation and rhythm features from audio, decompose into intonation fluctuation and rhythm pattern data, and generate a deviation vector;

[0021] If the intonation fluctuation component in the judgment vector exceeds the threshold, it is marked as prosodic deviation, and the prosodic adjustment index is obtained after quantification.

[0022] As an embodiment of the present invention, the steps for obtaining the standardized feature vector include:

[0023] The first feature set is generated by combining the preliminary speech feature set with the prosody adjustment parameters. If the dimension exceeds the threshold, it is reduced by principal component analysis; otherwise, it is directly used as the second feature set.

[0024] After modeling with a long short-term memory network, standardized feature vectors are obtained by standard deviation normalization.

[0025] As an embodiment of the present invention, the step of determining the degree of difference in cross-language comparison includes:

[0026] Extract speech features to generate an initial vector, adjust it if it is inconsistent with the dimension of the standard pronunciation vector, and then calculate the Euclidean distance;

[0027] The similarity or difference in pronunciation is determined by comparing the distance value with a threshold.

[0028] As one embodiment of the present invention, the steps for obtaining targeted feedback data include:

[0029] Determine if the cross-language difference exceeds a threshold, and extract personalized sub-vectors from the standardized feature vectors;

[0030] Based on the sub-vectors, an adjusted phoneme sequence is generated, the pitch transformation path is determined, and then targeted feedback data is generated.

[0031] As one embodiment of the present invention, the steps for obtaining the final analysis result include:

[0032] Extract time series data from the feedback data, and calculate the frequency distribution and the time series correlation between each parameter;

[0033] The system determines whether the correlation exceeds a threshold, fits the dynamic relationship, and analyzes the matching degree between intonation and pause rhythm to obtain the final result.

[0034] Another objective of this invention is to provide a foreign language learning reading quality analysis system based on voice interaction, characterized in that the system applies the above-mentioned method, including:

[0035] The audio acquisition module is used to collect audio recordings of learners reading aloud.

[0036] The feature extraction module, connected to the audio acquisition module, uses the Mel frequency cepstral coefficient algorithm to extract the fundamental frequency contour and spectral envelope features to obtain a preliminary speech feature set;

[0037] The audio alignment module, connected to the feature extraction module, uses a dynamic time warping algorithm to align the duration and energy distribution of the learner's audio with that of the standard audio, and determines the deviation vector.

[0038] The prosody analysis module, connected to the audio alignment module, extracts prosodic sub-features of intonation fluctuation and rhythm from the deviation vector, marks prosodic deviations based on the comparison results of intonation fluctuation and threshold, and obtains prosodic adjustment indicators.

[0039] The feature standardization module is connected to the prosody analysis module. It combines the prosody adjustment index with the preliminary speech feature set and processes it through a long short-term memory network to obtain a standardized feature vector.

[0040] The difference calculation module is connected to the feature standardization module. It calculates the Euclidean distance between the vector and the standard pronunciation vector to determine the cross-language comparison difference. Based on the comparison result of the difference and the threshold, it extracts personalized sub-vectors from the standardized feature vector and determines their corresponding prosodic rhythm adjustment paths.

[0041] The feedback generation module, connected to the difference calculation module, generates an adaptive improvement sequence based on the targeted feedback data to obtain the final reading quality analysis result.

[0042] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0043] This invention discloses a method for assessing and improving reading quality based on speech feature analysis and cross-language comparison. Addressing the challenges of prosodic deviations and cross-language pronunciation differences in learners' reading, the method collects learner audio signals and applies the Mel-frequency cepstral coefficient algorithm to extract fundamental frequency contours and spectral envelope features, constructing a preliminary speech feature set. A dynamic time warping algorithm is then used to align the duration and energy distribution of the learner's audio with that of a standard reference audio, generating a deviation vector. Prosodic sub-features, such as intonation fluctuations and rhythmic patterns, are extracted from the deviation vector to determine prosodic deviations and generate adjustment indicators. Combined with long short-term memory network processing, a standardized feature vector is generated. The cross-language comparison difference degree is determined by calculating the Euclidean distance to the standard pronunciation vector. Personalized sub-vectors are extracted from samples with high differences, generating targeted feedback data and adaptive improvement sequences. This invention, through the integration of prosodic adjustment and cross-language difference analysis, significantly improves the accuracy of reading quality assessment and the effectiveness of personalized improvement, providing efficient technical support for language learning. Attached Figure Description

[0044] Fig. 1 This is a flowchart of a foreign language learning reading quality analysis system and method based on voice interaction according to the present invention.

[0045] Fig. 2 This is a schematic diagram of a foreign language learning reading quality analysis system and method based on voice interaction according to the present invention.

[0046] Fig. 3 This is another schematic diagram of a foreign language learning reading quality analysis system and method based on voice interaction according to the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] like Figs. 1-3 This embodiment of a method for analyzing the quality of foreign language learning and reading aloud based on voice interaction may specifically include:

[0049] S101. By collecting the learner's reading audio signal and applying the Mel frequency cepstral coefficient algorithm to extract the fundamental frequency profile and spectral envelope features, a preliminary speech feature set is obtained.

[0050] The acquired learner's audio signal is converted from audio to frequency domain using short-time Fourier transform to generate a first audio time-frequency map. The Mel-frequency cepstral coefficient algorithm is then used to extract fundamental frequency contour data and spectral envelope data from the first audio time-frequency map, resulting in a first feature dataset. If the dimensionality of the first feature dataset exceeds a preset threshold, principal component analysis is used to reduce the dimensionality of the first feature dataset, resulting in a second feature dataset. Based on the time segment information of the second feature dataset, the fundamental frequency contour data and the spectral envelope data are integrated to generate the final speech feature set.

[0051] In this embodiment, audio is collected using a microphone with a sampling rate of 44.1kHz and quantization precision of 16 bits, and saved in WAV format. A short-time Fourier transform is implemented using the Librosa library in Python, with a window length of 20ms and a step size of 10ms, to generate the first audio time-frequency map. The Mel-frequency cepstral coefficients (MFCC) algorithm is used to extract 13-dimensional features (containing 12 coefficients + 1 energy value) to obtain the fundamental frequency profile (extracted with the aid of the YIN algorithm, frequency range 50-500Hz) and spectral envelope data. Dimensionality reduction is performed: a preset dimensionality threshold of 30 is used. When the dimensionality of the first feature dataset exceeds this limit, the PCA algorithm from the Scikit-learn library is used to retain 95% of the variance contribution for dimensionality reduction. The second feature dataset is divided into 100ms time segments, and the fundamental frequency profile and spectral envelope data are aligned using timestamps to generate the final speech feature set.

[0052] S102. Based on the preliminary speech feature set, the dynamic time warping algorithm is used to align the duration information and energy distribution of the learner's audio with the standard reference audio, and determine the alignment deviation vector.

[0053] Audio signals are acquired from learner audio and standard reference audio. Speech feature sets are extracted using short-time Fourier transform to obtain a first feature matrix containing duration information and energy distribution. A dynamic time warping algorithm is used to align the first feature matrix in time series, resulting in an aligned time mapping path. The differences in duration information and energy distribution between the learner audio and the standard reference audio are calculated based on the time mapping path to obtain a first deviation vector. If the matching accuracy of the first deviation vector is determined to be lower than a preset threshold, the first feature matrix is ​​adjusted using an interpolation method to generate a second feature matrix, and dynamic time warping is performed again to obtain an optimized second deviation vector.

[0054] In this embodiment, the duration (converted by frame counting, each frame corresponds to 10ms) and energy distribution (calculated by the short-time energy formula, with the energy threshold set to -40dB) are extracted from the learner and standard audio to generate a 100×2 dimension first feature matrix; the Dynamic Time Warping (DTW) algorithm (implemented based on the Librosa library) is used to calculate the time mapping path with a maximum time offset of 50 frames; the difference between duration and energy distribution is calculated using mean squared error (MSE) to generate a first deviation vector; the preset matching accuracy threshold is 0.85 (range 0-1). When the accuracy is insufficient, linear interpolation is used to extend the length of the feature matrix to 1.2 times that of the standard audio, and DTW is re-executed to obtain a second deviation vector.

[0055] S103. Obtain prosodic sub-features, including intonation fluctuation and rhythm pattern, from the aligned deviation vector. If the intonation fluctuation in the deviation vector exceeds a preset threshold, it is marked as a prosodic deviation, and the prosodic adjustment index is obtained.

[0056] Pitch and rhythm features are obtained from the input audio data using Fast Fourier Transform (FFT). The audio signal is then decomposed using time-frequency analysis to obtain pitch fluctuation data and rhythm pattern data. Based on the pitch fluctuation and rhythm pattern data, a deviation vector is generated using Principal Component Analysis (PCA). Vector analysis is then used to reduce the dimensionality of the pitch fluctuation and rhythm pattern data to obtain the deviation vector. If the pitch fluctuation component in the deviation vector exceeds a preset threshold, a prosodic deviation is determined through comparison. A threshold judgment method is used to analyze the deviation vector to obtain a prosodic deviation marker. Based on the prosodic deviation marker, an adjustment index is generated using a linear regression algorithm. Data processing methods are then used to quantify the prosodic deviation marker to obtain the prosodic adjustment index.

[0057] In this embodiment, intonation features in the 20-500Hz frequency band are extracted using Fast Fourier Transform (FFT), and rhythm features (beat interval range of 0.2-2 seconds) are extracted using autocorrelation. Wavelet transform (db4 wavelet basis) is used to decompose into intonation fluctuation data (represented by the fundamental frequency standard deviation) and rhythm pattern data (represented by the beat interval sequence). PCA is used to reduce the dimensionality of the intonation and rhythm data to a 2-dimensional vector, where the first component corresponds to intonation fluctuation and the second component corresponds to rhythm pattern. The intonation fluctuation threshold is set to 50Hz (based on native speaker pronunciation standard deviation statistics), and values ​​exceeding this threshold are marked as prosodic deviation. A linear regression algorithm (y = 0.5x + 0.2, where x is the deviation vector value) is used to convert the deviation markers into prosodic adjustment indicators in the range of 0-1 (1 indicates a significant adjustment is needed).

[0058] S104. A long short-term memory network is used to process the combination of prosodic adjustment index and preliminary speech feature set to obtain a standardized feature vector representation.

[0059] A preliminary speech feature set and prosody adjustment parameters are obtained from the input speech signal. The speech signal is then decomposed in the frequency domain using Fourier transform to obtain a spectral feature vector. This spectral feature vector is combined with the pitch and volume features from the prosody adjustment parameters to generate a first feature set. The dimensionality of the first feature set is then determined to be within a preset threshold. If it exceeds this threshold, principal component analysis is used to reduce the dimensionality of the first feature set, resulting in a second feature set. If the dimensionality does not exceed the threshold, the first feature set is directly used as the second feature set. A long short-term memory network is used to perform time series modeling on the second feature set, generating a sequence feature vector containing time dependencies, resulting in a third feature set. The third feature set is then standardized using standard deviation normalization to generate a standardized feature vector representation.

[0060] The initial speech feature set (13-dimensional MFCC) and the prosody adjustment index (1-dimensional) are concatenated to form a 14-dimensional first feature set. The pitch feature is taken from the fundamental frequency profile, and the volume feature is taken from the short-time energy. The preset dimension threshold is 20. If it exceeds 20, the dimension is reduced by PCA. Otherwise, it is directly retained. A 2-layer LSTM network (64 neurons per layer, dropout=0.2) is constructed to perform time series modeling on the feature set and output a 32-dimensional sequence feature vector. The feature vector is mapped to the interval [-1,1] by standard deviation normalization (x=(x-μ) / σ, μ is the mean, σ is the standard deviation).

[0061] S105. For the standardized feature vector representation, the cross-linguistic comparison difference is determined by calculating the Euclidean distance with the standard pronunciation vector.

[0062] Feature vectors are obtained from the input speech data. Speech features are extracted using a pre-established language model to generate an initial feature vector, resulting in the feature vector before standardization. If the dimension of the feature vector before standardization is inconsistent with the dimension of the standard pronunciation vector, a linear transformation is used to adjust the dimension, generating the standardized feature vector. The Euclidean distance formula is then used. Where n is the dimension of the feature vector, x i For the i-th component of the standardized eigenvector, y i The i-th component of the standard pronunciation vector is used, and the calculation result is retained to four decimal places. The distance between the standardized feature vector and the standard pronunciation vector is calculated to obtain the distance value. Based on the comparison of the distance value with a preset threshold, if the distance value is less than the threshold, the language pronunciation is determined to be similar; otherwise, a difference is determined, and the determination result of the cross-language comparison difference degree is obtained.

[0063] In this embodiment, the standard pronunciation vector is taken from a native speaker reading corpus (containing 1000+ samples) and generated through the same process as S101-S104. If the dimensions are inconsistent, the feature vector is projected to the standard dimension using a least squares linear transformation (y = Wx + b). The distance is calculated strictly according to the Euclidean distance formula, retaining 4 significant decimal places. The preset distance threshold is 5.0 (based on cross-language pronunciation difference statistics). If the distance is less than the threshold, it is considered similar; otherwise, it is considered that there is a difference (difference degree = distance / threshold, range 0-∞).

[0064] S106. If the cross-language comparison difference is higher than the preset threshold, then extract personalized sub-vectors from the standardized feature vectors, determine the prosodic rhythm adjustment path corresponding to these sub-vectors, and obtain targeted feedback data.

[0065] Language features are extracted from a multilingual text corpus, and semantic vectors are generated using a word embedding algorithm. If the Euclidean distance between the semantic vectors exceeds a preset threshold, the cross-lingual comparison difference is determined. Based on the cross-lingual comparison difference, a subset of features is extracted from the standardized feature vectors using principal component analysis to generate personalized sub-vectors, resulting in personalized weight allocation. Using these personalized sub-vectors, an adjusted phoneme sequence is generated using a phoneme sequence adjustment algorithm, and rhythm pattern matching results are obtained to determine the pitch transformation path. Targeted feedback data is generated based on the pitch transformation path, analysis results are generated through user behavior analysis, and the language feature extraction process is updated using a data annotation optimization algorithm to obtain the optimized semantic vectors.

[0066] In this embodiment, the multilingual corpus covers five languages: English, Japanese, Korean, Russian, and French.

[0067] Word2Vec generates 300-dimensional semantic vectors, with an Euclidean distance threshold of 10.0 (based on the language similarity matrix). PCA is used to extract the top three principal components (corresponding to pitch, speech rate, and pause features) from the standardized feature vectors as personalized sub-vectors. A dynamic programming algorithm (Dijkstra's algorithm) is employed to find the shortest adjustment path between the sub-vectors and the standard vectors; for example, the pitch adjustment path is "lower by 50Hz → hold for 2 seconds → fine-tune by 10Hz". Feedback data includes text prompts (such as "the pitch is too high, it is recommended to lower the pitch") and audio examples (based on TTS-synthesized and adjusted pronunciation). The semantic vector extraction weights are optimized using user click statistics.

[0068] S107. Generate an adaptive improvement sequence based on the targeted feedback data to obtain the final reading quality analysis results.

[0069] Temporal data of reading volume, speech rate, intonation, and pauses are extracted from targeted feedback data. The frequency distribution of this temporal data is calculated using Fourier transform to obtain a first frequency feature set. The Pearson correlation coefficient is used to calculate the temporal correlations between volume and speech rate, speech rate and intonation, pauses and intonation, and volume and pauses within the first frequency feature set, resulting in a first correlation matrix. If the correlation coefficients in the first correlation matrix are greater than a preset threshold, a linear regression algorithm is used to fit the dynamic relationships of volume, speech rate, intonation, and pauses, obtaining a first regression parameter set. Time series analysis tools are used to process the first regression parameter set, calculating the matching degree between intonation change trends and pause rhythms to obtain the final reading quality analysis results.

[0070] In this embodiment, the volume (dB value) and speech rate (syllables / ) are extracted from the feedback data every 100ms.

[0071] Data on intonation (fundamental frequency), pauses (duration > 200ms) were collected; Pearson correlation coefficient (range -1 to 1) was used to calculate the pairwise correlation between the four parameters, generating a 4×4 matrix; dynamic relationship fitting: the preset correlation coefficient threshold was 0.6, and if it exceeded, the parameter relationship was fitted by multiple linear regression (y = β0 + β1x1 + ... + β4x4); the matching degree between intonation and pauses was predicted by ARIMA time series model (range 0-100 points), and a final analysis report containing scores and improvement suggestions (such as "the pause in the third sentence is too short, it is recommended to extend it to 0.5 seconds") was generated.

[0072] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for analyzing the reading quality of foreign language learning based on voice interaction, characterized in that, The method includes: Audio recordings of learners reading aloud are collected, and the fundamental frequency profile and spectral envelope features are extracted using the Mel frequency cepstral coefficient algorithm to obtain a preliminary speech feature set; The dynamic time warping algorithm is used to align the duration and energy distribution of learner audio with that of standard audio to determine the bias vector; Extract intonation fluctuation and rhythmic sub-features from the deviation vector, and mark intonation fluctuation exceeding the threshold as rhythmic deviation to obtain rhythmic adjustment index; By combining prosodic adjustment indices and preliminary speech feature sets, standardized feature vectors are obtained through long short-term memory network processing. Calculate the Euclidean distance between this vector and the standard pronunciation vector to determine the degree of difference in cross-linguistic comparison; If the difference exceeds the threshold, extract personalized sub-vectors from the standardized feature vectors, determine the corresponding rhythm and cadence adjustment path, and obtain targeted feedback data. Based on the feedback data, an adaptive improvement sequence is generated, and the final reading quality analysis results are obtained.

2. The method according to claim 1, characterized in that, The steps for extracting a preliminary speech feature set include: The acquired audio was subjected to short-time Fourier transform to generate a time-frequency graph. The fundamental frequency profile and spectral envelope data were extracted using the Mel frequency cepstral coefficient algorithm to obtain the first feature dataset. If the dimensionality of the dataset exceeds the threshold, principal component analysis is used to reduce the dimensionality to obtain the second feature dataset, which is then integrated to generate the final speech feature set.

3. The method according to claim 1, characterized in that, The steps to determine the alignment deviation vector include: Feature matrices containing duration and energy distribution are extracted from learners and standard audio, and time series alignment is performed using a dynamic time warping algorithm to obtain a time mapping path. The first deviation vector is obtained based on the path calculation difference. It is determined that the matching accuracy is insufficient. After adjusting the feature matrix, the alignment is re-aligned to obtain the optimized deviation vector.

4. The method according to claim 1, characterized in that, The steps to obtain prosodic adjustment indicators include: Extract intonation and rhythm features from audio, decompose into intonation fluctuation and rhythm pattern data, and generate a deviation vector; If the intonation fluctuation component in the judgment vector exceeds the threshold, it is marked as prosodic deviation, and the prosodic adjustment index is obtained after quantification.

5. The method according to claim 1, characterized in that, The steps to obtain the standardized feature vectors include: The first feature set is generated by combining the preliminary speech feature set with the prosody adjustment parameters. If the dimension exceeds the threshold, it is reduced by principal component analysis; otherwise, it is directly used as the second feature set. After modeling with a long short-term memory network, standardized feature vectors are obtained by standard deviation normalization.

6. The method according to claim 1, characterized in that, The steps to determine cross-linguistic comparison differences include: Extract speech features to generate an initial vector, adjust it if it is inconsistent with the dimension of the standard pronunciation vector, and then calculate the Euclidean distance; The similarity or difference in pronunciation is determined by comparing the distance value with a threshold.

7. The method according to claim 1, characterized in that, The steps to obtain targeted feedback data include: Determine if the cross-language difference exceeds a threshold, and extract personalized sub-vectors from the standardized feature vectors; Based on the sub-vectors, an adjusted phoneme sequence is generated, the pitch transformation path is determined, and then targeted feedback data is generated.

8. The method according to claim 1, characterized in that, The steps to obtain the final analysis results include: Extract time series data from the feedback data, and calculate the frequency distribution and the time series correlation between each parameter; The system determines whether the correlation exceeds a threshold, fits the dynamic relationship, and analyzes the matching degree between intonation and pause rhythm to obtain the final result.

9. A foreign language learning reading quality analysis system based on voice interaction, characterized in that, The system applies the method according to any one of claims 1-8, including: The audio acquisition module is used to collect audio recordings of learners reading aloud. The feature extraction module, connected to the audio acquisition module, uses the Mel frequency cepstral coefficient algorithm to extract the fundamental frequency contour and spectral envelope features to obtain a preliminary speech feature set; The audio alignment module, connected to the feature extraction module, uses a dynamic time warping algorithm to align the duration and energy distribution of the learner's audio with that of the standard audio, and determines the deviation vector. The prosody analysis module, connected to the audio alignment module, extracts prosodic sub-features of intonation fluctuation and rhythm from the deviation vector, marks prosodic deviations based on the comparison results of intonation fluctuation and threshold, and obtains prosodic adjustment indicators. The feature standardization module is connected to the prosody analysis module. It combines the prosody adjustment index with the preliminary speech feature set and processes it through a long short-term memory network to obtain a standardized feature vector. The difference calculation module is connected to the feature standardization module. It calculates the Euclidean distance between the vector and the standard pronunciation vector to determine the cross-language comparison difference. Based on the comparison result of the difference and the threshold, it extracts personalized sub-vectors from the standardized feature vector and determines their corresponding prosodic rhythm adjustment paths. The feedback generation module, connected to the difference calculation module, generates an adaptive improvement sequence based on the targeted feedback data to obtain the final reading quality analysis result.

Images