A voice recognition method and system based on big data
By constructing a multi-dimensional matching index and a long short-term memory neural network model, combined with the AMPD peak detection algorithm, the problem of inaccurate tone recognition in traditional speech recognition technology is solved, and efficient and accurate speech recognition in complex scenarios is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU JIUSI INTELLIGENT TECH CO LTD
- Filing Date
- 2026-04-08
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional speech recognition technology suffers from low accuracy when processing Chinese tones, especially in complex scenarios where words with similar tones are easily confused. Furthermore, existing methods rely on manual annotation or fixed rules, resulting in low efficiency and an inability to accurately capture subtle differences and dynamic changes in tones.
By collecting speech signal frames and their fundamental frequencies, multiple matching indices (such as the first to fourth matching indices) are constructed to reflect tone characteristics. Combined with a long short-term memory artificial neural network model, feature extraction and recognition are performed. The AMPD peak detection algorithm is used to annotate key frames, thereby improving the accuracy and robustness of tone recognition.
It significantly improves the accuracy and robustness of Chinese speech recognition, enhances the adaptability to complex tone patterns, maintains high efficiency in noisy environments, and improves the accuracy of tone recognition and the stability of the system.
Smart Images
Figure CN121983036B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of speech recognition technology. More specifically, this invention relates to a speech recognition method and system based on big data. Background Technology
[0002] With the continuous advancement of artificial intelligence technology and the sustained improvement of computing power, speech recognition, as an important means of human-computer interaction, has been widely applied in various fields such as smartphones, smart speakers, in-vehicle systems, online customer service, education, and medical assistance. Traditional speech recognition systems mainly rely on the collaborative work of acoustic models, language models, and pronunciation dictionaries to convert speech into text through preprocessing, feature extraction, and pattern matching of the input speech signal. However, due to the influence of many factors such as language diversity, dialect differences, noise interference, and fluctuations in speech rate, the accuracy and robustness of traditional speech recognition technology in complex scenarios still face significant challenges.
[0003] Especially in Chinese speech recognition, Mandarin has four basic tones (high level, rising, falling-rising, and falling), and the same syllable can correspond to completely different Chinese characters depending on its tone. Tone information is crucial for accurately recognizing text content. However, in existing speech recognition technologies, tone is usually treated as a secondary attribute of speech, lacking in-depth and detailed modeling and differentiation, leading to confusion between words with similar tones. With the development of big data and machine learning technologies, feature learning and modeling based on large-scale speech data has become a trend. By statistically analyzing large amounts of real speech data, more representative speech feature patterns can be extracted, thereby improving recognition accuracy and system adaptability. Simultaneously, utilizing implicit information such as context and intonation trends in big data also provides new ideas for improving the accuracy of tone recognition.
[0004] However, traditional speech recognition methods have significant limitations when processing Chinese tones. Traditional methods usually rely on manually labeled tone data or preset fixed tone recognition rules, which not only increases the complexity of data preparation and model training, but also makes them inefficient when dealing with large-scale speech data or real-time speech stream processing. They often fail to accurately capture subtle differences and dynamic changes in tones, leading to tone recognition errors and consequently, low accuracy in speech content recognition. Summary of the Invention
[0005] To address the problem of low accuracy in speech content recognition mentioned in the background art, the present invention provides solutions in the following aspects.
[0006] In a first aspect, the present invention provides a speech recognition method based on big data, comprising: acquiring each speech signal frame and its fundamental frequency within a target region; recording the sequence of fundamental frequencies of all speech signal frames within a set neighborhood radius centered on each speech signal frame as a target fundamental frequency sequence, and performing first-order difference processing on the target fundamental frequency sequence to obtain a difference sequence; obtaining a first matching index between each speech signal frame and the rising tone, and a second matching index between each speech signal frame and the departing tone; wherein the first matching index shows an upward trend with the target fundamental frequency sequence. The sum of all positive values in the difference sequence is positively correlated with the absolute value of the sum of all negative values in the difference sequence; the second matching index shows a decreasing trend with the target fundamental frequency sequence. The absolute values of the sum of all negative numbers in the difference sequence are positively correlated, while the sum of all positive numbers in the difference sequence is negatively correlated. The third matching index of each speech signal frame with the first tone and the fourth matching index with the third tone are obtained. The third and fourth matching indices characterize the fluctuation level of each speech signal frame. The maximum values of the first, second, third, and fourth matching indices corresponding to each speech signal frame are used as the feature values of each speech signal frame, and the labels of each feature value are obtained. The labels of each speech signal frame and its feature values are input into a trained machine learning model to obtain the speech content of each speech signal frame.
[0007] The above technical solution collects speech signal frames and their fundamental frequencies, and extracts the fundamental frequency sequence and its differential features within a set neighborhood. It constructs matching indices that reflect intonation features from multiple perspectives, which are used to identify the trend features and fluctuation characteristics of different tones. This effectively enhances the ability to distinguish speech tone patterns. The maximum value among multiple matching indices is used as the feature value and combined with label information to train the model, enabling more accurate identification of the speech content corresponding to the speech signal frame. This improves the accuracy and robustness of speech recognition in tone-sensitive scenarios.
[0008] Furthermore, the first matching index for: , For the first An upward trend exists in the target fundamental frequency sequence of each speech signal frame. value, For the first The sum of all positive numbers in the difference sequence of each speech signal frame. For the first The absolute value of the sum of all negative numbers in the differential sequence of several speech signal frames.
[0009] The above technical solution constructs an index to characterize the degree of matching between a speech signal frame and the rising tone. The design comprehensively considers the significance of the rising trend in the fundamental frequency sequence and the proportion of the rising change in the differential sequence relative to the overall change, which helps to enhance the sensitivity of capturing the rising tone features.
[0010] Furthermore, the second matching index for: , For the first There is a decreasing trend in the target fundamental frequency sequence of each speech signal frame. value, For the first The sum of all positive numbers in the difference sequence of each speech signal frame. For the first The absolute value of the sum of all negative numbers in the differential sequence of several speech signal frames.
[0011] The above technical solution constructs an index to measure the degree of matching between the speech signal frame and the tone drop, and combines the significance of the downward trend in the fundamental frequency sequence with the proportion of the overall downward amplitude. This can effectively enhance the sensitivity to tone drop features, not only highlighting the key features of the tone drop in speech fundamental frequency changes, but also reducing the interference of other changing factors through normalization processing, thereby improving the accuracy and discriminativeness of tone recognition.
[0012] Furthermore, the third matching index for: , For the first The mean of all fundamental frequencies in the target fundamental frequency sequence of a speech signal frame. For the first The standard deviation of all fundamental frequencies in the target fundamental frequency sequence of a speech signal frame.
[0013] The above technical solution constructs an index to measure the relationship between the stationarity and volatility of speech signal frames. By comparing the overall average level of the fundamental frequency sequence with its volatility, it helps to highlight speech segments with relatively stable fundamental frequency changes and small volatility amplitudes, thereby effectively characterizing the features of tone types such as the first tone.
[0014] Furthermore, the fourth matching index is specifically:
[0015] The sequence consisting of the fundamental frequencies of all speech signal frames within a predetermined neighborhood radius to the left of each speech signal frame is denoted as the first sequence, and the sequence consisting of the fundamental frequencies of all speech signal frames within a predetermined neighborhood radius to the right of each speech signal frame is denoted as the second sequence; the fourth matching index for: , For the first A downward trend exists in the first sequence of the speech signal frames. value, For the first An upward trend exists in the second sequence of the speech signal frames. value, These are the preset hyperparameters.
[0016] The above technical solution constructs an index to measure the symmetry and continuity of fundamental frequency trend changes by comparing the trend differences of fundamental frequency sequences in the left and right neighborhoods of a speech signal frame. This helps to identify frames with obvious inflection points or fluctuations in tone, and is particularly suitable for tone recognition with complex tonal changes, such as the third tone.
[0017] Furthermore, the step of obtaining the label of each feature value specifically involves: using each feature value as input to the AMPD peak detection algorithm to obtain the peak value of each feature value, recording the label of the feature value of the speech signal frame corresponding to the peak value as 1, and recording the label of the feature value of the speech signal frame corresponding to the non-peak value as 0.
[0018] Furthermore, the machine learning model is a long short-term memory artificial neural network model.
[0019] Furthermore, it also includes training the machine learning model, specifically: inputting the training set into the pre-built machine learning model for training; during the training process, calculating the loss between the output predicted value and the label; adjusting the model parameters using gradient descent to minimize the prediction error; iteratively adjusting the parameters of the machine learning model until the loss is less than a certain value or the set number of training iterations is reached, and finally obtaining the trained machine learning model.
[0020] Furthermore, each speech signal frame is subjected to noise reduction processing.
[0021] In a second aspect, the present invention provides a speech recognition system based on big data, including a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the speech recognition method based on big data described above is implemented.
[0022] The beneficial effects of this invention are as follows:
[0023] This invention constructs multiple matching indices that reflect fundamental frequency variation characteristics, comprehensively extracts key variation trends and fluctuation characteristics of different tone types in speech signals, accurately labels feature frames by combining peak detection algorithms, and achieves efficient learning and recognition with the help of neural network models. While improving tone recognition accuracy, it also enhances the adaptability to complex tone patterns and robustness to noisy environments, significantly optimizing the accuracy of speech recognition. Attached Figure Description
[0024] Figure 1 This is a schematic flowchart illustrating a speech recognition method based on big data according to an embodiment of the present invention;
[0025] Figure 2 This is a schematic diagram illustrating the structure of a big data-based speech recognition system according to an embodiment of the present invention. Detailed Implementation
[0026] An embodiment of a speech recognition method based on big data.
[0027] like Figure 1 As shown, a fire early warning flowchart based on multi-detector data fusion according to an embodiment of the present invention includes the following steps:
[0028] S1: Collect each voice signal frame and its fundamental frequency within the target area.
[0029] In one embodiment, voice data within a target area can be collected using recording devices such as microphones. Since the raw voice data may contain environmental noise that could interfere with subsequent speech recognition, Wiener filtering is applied to the collected voice data for noise reduction to improve the clarity of the voice signal.
[0030] After denoising, to accurately locate speech segments and avoid misidentification due to pauses in the speaker, the VAD endpoint detection algorithm is used to analyze the denoised speech data and identify the effective regions actually containing speech. Because speech signals are non-stationary, they need to be segmented into short frames during processing to maintain relatively stable speech signal characteristics within each frame for subsequent feature extraction and analysis. In this embodiment, each frame of speech signal is set to 25 milliseconds, and the interval between frames is 10 milliseconds. Specific values can be adjusted according to different needs.
[0031] Furthermore, to reduce the adverse effects of spectrum leakage, a Hamming window is applied to each frame of the speech signal after framing, thereby enhancing the accuracy of spectrum analysis. After the above processing, the preprocessed speech dataset is obtained, which contains speech signal frames at different time points within the target region. In this embodiment, the fundamental frequency information corresponding to each frame of the speech signal is further extracted, thereby providing support for subsequent speech feature extraction and analysis.
[0032] S2: Obtain the first matching index of each speech signal frame with the rising tone, the second matching index with the falling tone, the third matching index with the high level tone, and the fourth matching index with the rising tone.
[0033] In one embodiment, the first matching index for: , For the first An upward trend exists in the target fundamental frequency sequence of each speech signal frame. value, For the first The sum of all positive numbers in the difference sequence of each speech signal frame. For the first The absolute value of the sum of all negative numbers in the differential sequence of several speech signal frames.
[0034] The above scheme effectively characterizes the smooth rising characteristic of the second tone in speech by weighting and combining the significance of the fundamental frequency rise trend with the total amount of positive change in the speech signal frame. By quantifying the intensity of the rise trend in the speech signal and combining it with its actual change amplitude, the accurate extraction of the rising tone feature is achieved.
[0035] Second matching index for: , For the first There is a decreasing trend in the target fundamental frequency sequence of each speech signal frame. value, For the first The sum of all positive numbers in the difference sequence of each speech signal frame. For the first The absolute value of the sum of all negative numbers in the differential sequence of several speech signal frames.
[0036] The above scheme effectively characterizes the rapid descent of vowel tones in speech by weighting the descent trend of the fundamental frequency in the speech signal frame with the magnitude of the negative change. By quantifying the significance of the descent trend and combining it with the actual amplitude of the signal change, speech segments with vowel tones can be identified more accurately.
[0037] The third matching index for: , For the first The mean of all fundamental frequencies in the target fundamental frequency sequence of a speech signal frame. For the first The standard deviation of all fundamental frequencies in the target fundamental frequency sequence of a speech signal frame.
[0038] The above scheme calculates the ratio of the mean fundamental frequency to its fluctuation in a speech signal frame, effectively reflecting the stability characteristics of the fundamental frequency in speech, thus capturing the relatively stable and minimally changing fundamental frequency of the level tone during pronunciation. When discriminating tones in speech signals, this scheme highlights the tone pattern corresponding to a stable fundamental frequency, improving the accuracy of level tone recognition. Especially in real-world speech environments with fluctuating speech rate or noise interference, it maintains high discrimination ability and system robustness.
[0039] The fourth matching index is specifically defined as follows: the sequence composed of the fundamental frequencies of all speech signal frames within a set neighborhood radius on the left side of each speech signal frame is denoted as the first sequence, and the sequence composed of the fundamental frequencies of all speech signal frames within a set neighborhood radius on the right side is denoted as the second sequence.
[0040] Fourth Matching Index for: , For the first A downward trend exists in the first sequence of the speech signal frames. value, For the first An upward trend exists in the second sequence of the speech signal frames. value, These are the preset hyperparameters.
[0041] The above scheme effectively captures the typical fundamental frequency pattern of the third tone (rising tone) by comprehensively considering the trend characteristics of fundamental frequency changes in the neighborhoods on both sides of the speech signal frame. By jointly measuring the significance of the falling and rising trends and introducing a harmonic term to balance the difference between them, this method can enhance the sensitivity to complex fundamental frequency trends, thereby improving the ability to recognize the third tone.
[0042] S3: Take the maximum value of multiple matching indices of each speech signal frame as the feature value of each speech signal frame, and obtain the label of each feature value.
[0043] In one embodiment, obtaining the label of each feature value specifically involves: using each feature value as input to the AMPD peak detection algorithm to obtain the peak value of each feature value, labeling the feature value of the speech signal frame corresponding to the peak value as 1, and labeling the feature value of the speech signal frame corresponding to the non-peak value as 0.
[0044] By introducing the AMPD peak detection algorithm to analyze feature values, key frames with significant changes or representative features in speech signals can be effectively identified and assigned different labels, achieving ordered annotation of speech features. Utilizing the characteristic that peaks in temporal features typically correspond to important turning points in speech, the accuracy and representativeness of label assignment are improved. This helps subsequent models learn significant change patterns in speech signals more accurately, thereby enhancing the effectiveness of speech analysis and tone recognition.
[0045] S4: Input the labels of each speech signal frame and its feature values into the trained machine learning model to obtain the speech content of each speech signal frame.
[0046] In one embodiment, the machine learning model is a long short-term memory artificial neural network model. The method further includes training the machine learning model, taking the long short-term memory artificial neural network model as an example, specifically: inputting the training set into the pre-constructed long short-term memory artificial neural network model for training; during training, calculating the loss between the output predicted value and the label; adjusting the model parameters using gradient descent to minimize the prediction error; iteratively adjusting the parameters of the long short-term memory artificial neural network model until the loss is less than a certain value or a set number of training iterations is reached, ultimately obtaining the trained long short-term memory artificial neural network model.
[0047] The labels of each speech signal frame and its feature values are input into a trained long short-term memory artificial neural network model to obtain the speech content of each speech signal frame.
[0048] The present invention constructs a multi-dimensional matching index to finely characterize the matching relationship between each speech signal frame and different tones. Combining features such as fundamental frequency trend, fluctuation degree, and neighborhood symmetry, it can effectively enhance the sensitivity and recognition accuracy of speech tone changes. At the same time, it introduces a peak detection algorithm to accurately extract key speech frames, and uses a long short-term memory neural network model to model the mapping relationship between features and speech content, which is conducive to improving the model's learning ability and generalization ability of temporal information. Ultimately, it achieves efficient and accurate recognition of complex speech data, and has significant robustness and practical value.
[0049] An example of a speech recognition system based on big data:
[0050] like Figure 2 As shown, a block diagram of a speech recognition system based on big data according to an embodiment of the present invention includes a processor and a memory.
[0051] This invention also provides a speech recognition system based on big data. For example... Figure 2As shown, the system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement a big data-based speech recognition method according to the present invention.
[0052] The aforementioned big data-based speech recognition system also includes other components well-known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.
[0053] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented by computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.
[0054] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise explicitly specified.
[0055] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.
Claims
1. A speech recognition method based on big data, characterized in that, include: Collect each speech signal frame and its fundamental frequency within the target area; denote the sequence of fundamental frequencies of all speech signal frames within a set neighborhood radius centered on each speech signal frame as the target fundamental frequency sequence, and perform first-order differential processing on the target fundamental frequency sequence to obtain a differential sequence; Obtain the first matching index of each speech signal frame with the rising tone and the second matching index with the departing tone; the first matching index for: , For the first An upward trend exists in the target fundamental frequency sequence of each speech signal frame. value, For the first The sum of all positive numbers in the difference sequence of each speech signal frame. For the first The absolute value of the sum of all negative numbers in the difference sequence of each speech signal frame; the second matching index. for: , For the first There is a decreasing trend in the target fundamental frequency sequence of each speech signal frame. value; Obtain the third matching index of each speech signal frame with the first tone (yinping) and the fourth matching index with the third tone (shangsheng); the third and fourth matching indices characterize the fluctuation degree of each speech signal frame; the third matching index... for: , For the first The mean of all fundamental frequencies in the target fundamental frequency sequence of a speech signal frame. For the first The standard deviation of all fundamental frequencies in the target fundamental frequency sequence of a speech signal frame; The fourth matching index is specifically defined as follows: the sequence consisting of the fundamental frequencies of all speech signal frames within a predetermined neighborhood radius to the left of each speech signal frame is denoted as the first sequence, and the sequence consisting of the fundamental frequencies of all speech signal frames within a predetermined neighborhood radius to the right of each speech signal frame is denoted as the second sequence; the fourth matching index for: , For the first A downward trend exists in the first sequence of the speech signal frames. value, For the first An upward trend exists in the second sequence of the speech signal frames. value, These are preset hyperparameters; The maximum values of the first matching index, the second matching index, the third matching index, and the fourth matching index corresponding to each speech signal frame are used as the feature values of each speech signal frame, and the labels of each feature value are obtained. The speech signal frames and the labels of their feature values are then input into the trained machine learning model to obtain the speech content of each speech signal frame.
2. The speech recognition method based on big data according to claim 1, characterized in that, The specific steps for obtaining the labels for each feature value are as follows: Each feature value is used as input to the AMPD peak detection algorithm to obtain the peak value of each feature value. The feature value of the speech signal frame corresponding to the peak value is marked as 1, and the feature value of the speech signal frame corresponding to the non-peak value is marked as 0.
3. The speech recognition method based on big data according to claim 1, characterized in that, The machine learning model is a long short-term memory artificial neural network model.
4. The speech recognition method based on big data according to claim 1, characterized in that, This also includes training machine learning models, specifically: The training set is fed into a pre-built machine learning model for training. During the training process, the loss between the output prediction value and the label is calculated. Use gradient descent to adjust model parameters to minimize prediction error; The parameters of the machine learning model are iteratively adjusted until the loss is less than a set value or the set number of training iterations are reached, and finally a well-trained machine learning model is obtained.
5. The speech recognition method based on big data according to claim 1, characterized in that, The speech signal frames are then denoised.
6. A speech recognition system based on big data, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, it implements the speech recognition method based on big data as described in any one of claims 1 to 5.