Two-stage cascaded speech quality filtering method and system based on multi-feature weighted score

Abstract

The application discloses a two-stage cascaded speech quality filtering method and system based on multi-feature weighted scoring, and belongs to the field of speech signal processing. Through a data-driven feature screening method with difference ratio as the core, a two-stage cascaded architecture of energy pre-screening and multi-feature weighted scoring is constructed in the front end of the speech recognition engine: the first stage performs three-class shunting based on the root mean square energy double threshold, the threshold is determined by the positive and negative sample statistical quantile point, and only the boundary sample enters the second stage; the second stage extracts three types of orthogonal acoustic features of periodicity strength, fundamental frequency band energy proportion and low frequency energy proportion, and after segmenting and grading mapping, the features are weighted and scored according to the feature difference ratio, and the final decision result is output. The application can realize high-precision filtering with a computing cost close to that of traditional VAD, has interpretability and cross-scene migration ability, and can effectively reduce invalid recognition and save computing resources.

Description

Technical Field

[0001] This invention belongs to the field of speech signal processing technology, specifically relating to a speech quality filtering method and system deployed at the front end of a speech recognition engine, which can be applied to speech interaction scenarios in complex acoustic environments such as offices, factory workshops, and public places. Background Technology

[0002] With the widespread adoption of Automatic Speech Recognition (ASR) technology, voice interaction systems have been widely deployed in scenarios such as intelligent customer service, in-vehicle interaction, industrial inspection, and meeting recording. However, in practical applications, a large amount of invalid audio without effective semantics is sent to the recognition engine, triggering the entire recognition process, but the result is only meaningless output, forming the problem of invalid recognition. This has become a key bottleneck restricting the practical application of voice interaction systems.

[0003] Based on statistics from actual deployment scenarios, invalid recognition mainly comes from four types of audio: First, environmental noise triggers, such as air conditioner operation, mechanical equipment noise, background music, footsteps and other non-speech sound sources; second, low-quality audio interference, such as current microphone noise, transmission attenuation and blurred audio, equipment howling and so on; third, non-target speech mixed in, such as long-distance conversations in the background, TV broadcasts, mobile phone speakers and other non-target human voices; and fourth, silent and extremely weak noise segments, caused by insufficient accuracy of speech activity detection segmentation.

[0004] Feeding invalid audio into the speech recognition model can cause three problems: first, it generates incorrectly recognized text, polluting downstream business logic; second, it consumes recognition computing power, resulting in significant resource waste on edge devices or high-concurrency services; and third, it reduces system availability and user trust. Therefore, adding an efficient and accurate speech quality filtering stage to the front end of the speech recognition engine has significant engineering value.

[0005] However, existing technologies are insufficient for this speech quality filtering task, exhibiting the following shortcomings: Traditional Voice Activity Detection (VAD) relies solely on short-time energy or simple spectral features to determine the presence of sound, failing to distinguish between sound and valid human voices. It is prone to missing low-energy valid human voices and misjudging high-energy device noise. A single fixed threshold filtering scheme faces a dilemma between accuracy and recall: an excessively high threshold may mistakenly delete low-volume valid speech, reducing recall; an excessively low threshold may allow high-energy noise to pass through, resulting in a high false trigger rate and an inability to adapt to varying acoustic environments. End-to-end deep learning judgment... While other solutions offer high accuracy, they suffer from large model parameters, high inference latency, and stringent requirements for computing power and memory, making them difficult to deploy on resource-constrained edge devices. Furthermore, the discrimination process is a black box, making it impossible to perform interpretable optimizations for specific scenarios. Single-acoustic feature discrimination solutions have poor robustness, as relying on a single feature cannot cover a wide variety of invalid audio types, and features effective for one type of noise often fail for another. Existing rule-based filtering solutions rely heavily on subjective selection of acoustic features based on engineers' experience, lacking quantifiable and reproducible screening criteria. When changing deployment scenarios, trial and error are required, resulting in poor transferability.

[0006] Therefore, existing technologies lack a lightweight speech quality filtering solution that can accurately distinguish between valid human voices and various invalid audio types, while maintaining extremely low computational overhead and possessing interpretability and transferability. Summary of the Invention

[0007] The purpose of this invention is to provide a two-level cascaded speech quality filtering method and system based on multi-feature weighted scoring, which achieves high-precision invalid audio filtering with low computational overhead, while ensuring the interpretability and cross-scenario transferability of the solution.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0009] A two-stage cascaded speech quality filtering method based on multi-feature weighted scoring includes the following steps:

[0010] S1. Perform sampling rate conversion and amplitude normalization preprocessing on the acquired raw audio stream to obtain a standardized audio signal;

[0011] S2. For the standardized audio signal, based on short-time energy and zero-crossing rate to detect the speech activity interval, the continuous audio stream is divided into several independent audio segments;

[0012] S3. Calculate the root mean square energy value for each audio segment, and perform a three-classification decision based on the root mean square energy value and a preset first threshold and second threshold: audio segments with root mean square energy values ​​lower than the first threshold are determined as invalid audio and are directly filtered; audio segments with root mean square energy values ​​higher than the second threshold are determined as valid speech and are directly passed; audio segments with root mean square energy values ​​between the first threshold and the second threshold are determined as boundary samples; wherein the first threshold is less than the second threshold.

[0013] S4. Extract preset multidimensional human voice-specific acoustic features from the boundary samples to obtain the corresponding feature value set;

[0014] S5. Perform segmented level mapping on each feature value in the feature value set, then sum the features according to the preset weights to obtain the comprehensive score, and compare the comprehensive score with the scoring threshold to obtain the final judgment result.

[0015] S6. Summarize the judgment results of all audio segments, output the audio segments judged as valid speech to the downstream speech recognition engine, and discard the audio segments judged as invalid audio.

[0016] Furthermore, the first threshold and the second threshold in step S3 are predetermined by the statistical distribution of the labeled dataset. Specifically, a labeled dataset including a set of valid positive speech samples and a set of invalid negative audio samples is constructed. The high quantile of the root mean square energy distribution of the negative sample set is taken as the first threshold, and the low quantile of the root mean square energy distribution of the positive sample set is taken as the second threshold. The interval between the two thresholds is the overlapping and undeterminable region of the positive and negative sample energy distributions.

[0017] Furthermore, the multidimensional human voice-specific acoustic features in step S4 include a first acoustic feature, a second acoustic feature, and a third acoustic feature: the first acoustic feature is the periodic intensity feature of the audio segment, which is obtained by periodically detecting the audio segment frame by frame and statistically analyzing the detection results, and is used to characterize the periodic characteristics of the vocal cord vibration of the human voice; the second acoustic feature is the fundamental frequency band energy ratio feature, which is the ratio of the energy of the fundamental frequency band of the human voice in the audio segment to the total energy of the segment, and is used to characterize the degree of concentration of human voice energy in the fundamental frequency band; the third acoustic feature is the low-frequency energy ratio feature, which is the ratio of the energy of the preset low-frequency band in the audio segment to the total energy of the segment, and is used to characterize the low-frequency energy distribution characteristics of the audio.

[0018] Furthermore, step S5 specifically includes: inputting the feature values ​​of each acoustic feature into the corresponding piecewise step-type scoring mapping function to map and obtain a confidence level with a value range of [0, 1]. The higher the confidence level value, the more the feature supports the determination that the current audio is a valid human voice; assigning a corresponding weight to each acoustic feature, the weight being positively correlated with the feature's ability to distinguish between positive and negative samples; and multiplying the confidence level of each acoustic feature by its corresponding weight and summing the results to obtain the comprehensive score of the audio segment.

[0019] Furthermore, the multidimensional human voice-specific acoustic features in step S4 are pre-determined using a data-driven screening method. The screening process includes: collecting diverse audio samples from the target scene, manually labeling them as positive and negative samples, and constructing a labeled dataset; extracting five categories of candidate acoustic features—time domain, frequency domain, cepstral, energy, and pitch—from all samples in the labeled dataset to form a candidate feature pool; calculating the difference ratio of each candidate feature on the positive and negative sample sets, where the difference ratio is the ratio of the absolute difference between the mean values ​​of the positive and negative samples to the joint standard error of the estimated mean values ​​of the two samples, used to quantify the feature's ability to distinguish between valid human voices and invalid audio; arranging all candidate features in descending order of difference ratio to construct a feature importance ranking table; and selecting the optimal feature subset from the feature importance ranking table based on four screening criteria: high discriminative power, cross-scene stability, low redundancy, and computational efficiency, as the multidimensional human voice-specific acoustic features.

[0020] A two-level cascaded speech quality filtering system based on multi-feature weighted scoring includes a first-level filtering module, a second-level filtering module, and a result output module;

[0021] The first-level filtering module is used to receive the original audio stream, complete preprocessing and speech activity segmentation, calculate the root mean square energy value of each audio segment, and perform a dual-threshold three-class classification decision after passing the preset first threshold and second threshold. The audio segments judged as invalid audio and the discard flag are output to the result output module, the audio segments judged as valid speech are directly output to the result output module, and the audio segments judged as boundary samples are output to the second-level filtering module.

[0022] The second-level filtering module is used to receive the boundary samples, extract multi-dimensional human voice-specific acoustic features, and calculate the comprehensive score by weighting the scores according to preset weights after segmentation and level mapping. After comparing with the scoring threshold, the judgment result and the corresponding audio segment are output to the result output module.

[0023] The result output module is used to receive audio segments and corresponding judgment results output by the first-level filtering module and the second-level filtering module, and after summarizing, output the valid speech to the downstream speech recognition engine and discard the invalid audio.

[0024] Furthermore, the first-level filtering module includes a VAD segmentation unit, an energy calculation unit, and an energy splitting unit. The VAD segmentation unit receives the original audio stream, performs sampling rate conversion and amplitude normalization preprocessing, detects speech activity intervals based on short-time energy and zero-crossing rate, segments the audio into several independent audio segments, and outputs them to the energy calculation unit. The energy calculation unit receives each audio segment, calculates the root mean square energy value of each audio segment, and outputs it to the energy splitting unit. The energy splitting unit receives the root mean square energy value of each audio segment, compares it with a preset first threshold and a second threshold, performs three-class splitting, outputs invalid audio and discard flags, and valid speech segments to the result output module, and outputs boundary samples to the second-level filtering module.

[0025] Furthermore, the second-level filtering module includes a feature extraction unit, a scoring calculation unit, and a threshold determination unit. The feature extraction unit receives boundary samples and extracts three types of acoustic features—periodic strength features, fundamental frequency band energy proportion features, and low-frequency energy proportion features—according to the frequency band range corresponding to each acoustic feature. The corresponding feature values ​​are then output to the scoring calculation unit. The scoring calculation unit receives the feature values ​​of the three types of acoustic features, performs segmented level mapping on each feature value using a scoring mapping function containing segmented thresholds, and then sums the values ​​according to preset weights to obtain a comprehensive score, which is then output to the threshold determination unit. The threshold determination unit receives the comprehensive score, compares it with the scoring threshold, and outputs the result of determining whether the corresponding audio segment is valid or invalid audio to the result output module.

[0026] Furthermore, the preset weights of each acoustic feature in the scoring calculation unit are positively correlated with the difference ratio of the corresponding feature; the difference ratio is the ratio of the absolute difference of the mean of the feature on the positive sample set and the negative sample set to the joint standard error.

[0027] Furthermore, the system is deployed at the front end of the speech recognition engine; the first threshold, the second threshold, the frequency band range of each acoustic feature, the segmented threshold of the scoring mapping function, the preset weight, and the scoring threshold are all configurable parameters that can be adjusted according to the deployment scenario.

[0028] Compared with the prior art, the present invention has the following detailed beneficial effects:

[0029] This invention employs a two-stage cascaded architecture of energy pre-screening and multi-feature weighted scoring. The first stage uses root mean square energy as a dual threshold to achieve three-class classification. The threshold boundary is determined by the statistical quantiles of the labeled dataset, directly classifying samples with clear energy at near-zero cost. Only boundary samples where positive and negative sample distributions overlap are transferred to the second stage for fine-grained discrimination. This architecture allows the vast majority of samples to be processed in the first stage with extremely low computational cost of O(N), with only a small number of boundary samples bearing the computational burden of feature extraction and scoring. Ultimately, the overall computational cost is close to that of traditional VAD schemes, while the discrimination accuracy is close to that of multi-feature fine-grained discrimination schemes. This fundamentally resolves the dilemma of accuracy and recall in single-threshold schemes, achieving a synergistic effect of obtaining fine-grained discrimination accuracy at the cost of a coarse-grained discriminator.

[0030] This invention selects three types of acoustic features from different dimensions in the time and frequency domains, whose physical mechanisms are orthogonal and whose failure modes do not overlap: periodic features characterize the vocal cord vibration characteristics of human voice in the time domain; fundamental frequency band energy proportion features characterize the energy distribution of human voice in the low frequency range; and low-frequency energy distribution features characterize the proportion of low frequencies in the overall audio in the frequency domain. These three types of features are effective against different types of invalid audio: periodic features are good at filtering non-periodic noise; fundamental frequency proportion features are good at filtering broadband noise; and low-frequency proportion features can effectively identify high-frequency howling and current microphone interference. By weighted summation and fusion of the discrimination results of the three types of features, the occasional failure of a single feature can be compensated by the other features, and any invalid audio type is captured by at least one type of feature. The overall robustness is significantly higher than that of single-feature schemes and arbitrarily stacked multi-feature schemes.

[0031] This invention proposes a feature selection methodology based on the difference ratio as the core indicator, supplemented by four selection criteria, upgrading feature selection from experience-based trial and error to data-driven quantitative decision-making. The difference ratio, calculated as the ratio of the mean difference between positive and negative samples to the distribution dispersion, quantifies the ability of a single feature to distinguish between valid human voices and invalid audio, possessing clear statistical significance. Combined with the four criteria of high discriminative power, cross-scene stability, low redundancy, and computational efficiency, it can select a statistically near-optimal subset of features. When changes in deployment scenarios lead to altered noise distribution, only new scene data needs to be collected and the selection process re-executed. Within hours, feature combinations and weights adapted to the new scene can be automatically exported, eliminating the need to redesign algorithms or train models. The migration cycle is shortened from redevelopment to simply rerunning the selection process, demonstrating extremely strong scene adaptability.

[0032] The entire discrimination chain of this invention uses white-box logic. From the root mean square energy value, various acoustic feature values ​​and feature confidence levels to the comprehensive score, all intermediate parameters can be output and traced. When misjudgment occurs in the field, maintenance personnel can directly locate the anomaly in the score of a specific feature and adjust the segment threshold or weight of that feature accordingly, without re-labeling data and retraining the model. Simultaneously, the solution has low overall computational cost, no complex iterative calculations, and millisecond-level latency for single-segment processing. It can be directly deployed on resource-constrained edge devices, possessing excellent field adaptability and maintainability. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of the speech quality filtering method of the present invention;

[0034] Figure 2 This is a flowchart of the data-driven key acoustic feature screening methodology of this invention;

[0035] Figure 3 This is a structural block diagram of the voice quality filtering system of the present invention. Detailed Implementation

[0036] To facilitate understanding of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the embodiments described are merely illustrative of the invention and should not be considered as specific limitations thereof.

[0037] Example 1:

[0038] This embodiment details the complete execution flow of the two-stage cascaded speech quality filtering method of the present invention. The overall flow is as follows: Figure 1 As shown, the specific steps are as follows:

[0039] Step S1: Audio signal acquisition and preprocessing

[0040] It receives raw audio streams from microphones, microphone arrays, or other acquisition devices, performs sampling rate conversion and amplitude normalization processing, unifies the audio from different acquisition devices to a standard format (in this embodiment, the unified sampling rate is 16kHz), eliminates dimensional differences between acquisition devices, and outputs a standardized audio signal.

[0041] Step S2: Speech activity detection and segmentation

[0042] The standardized audio signal output in step S1 is used to detect the speech activity range based on short-time energy and zero-crossing rate. The continuous audio stream is then divided into several independent audio segments, each representing a potential speech event. The segment length is typically between 0.5 seconds and several seconds.

[0043] Let the signal of the t-th frame be x. t(n), and the frame length is L, then the short-term energy E(t) and the zero-crossing rate ZCR(t) are respectively:

[0044] ;

[0045] ;

[0046] wherein, the short-term energy E(t) reflects the energy intensity of the signal within the frame, and the zero-crossing rate ZCR(t) reflects the frequency characteristic of the signal; noise and unvoiced sounds have a high zero-crossing rate, while voiced sounds have a low zero-crossing rate, and the combination of the two can preliminarily locate the voice activity interval.

[0047] Step S3: First-stage energy pre-screening and three-classification shunting

[0048] Each audio segment segmented in step S2 is received, and the root mean square (RMS) value of each audio segment is calculated as the energy representation:

[0049] ;

[0050] wherein, x(n) is the amplitude of the nth sampling point of the segment, and N is the total number of sampling points of the segment; this calculation only requires one traversal, with a time complexity of O(N).

[0051] The RMS value of each audio segment is compared with a preset first threshold T1 and a preset second threshold T2 (T1<T2) to perform three-classification decision:

[0052] ;

[0053] The decision rules are: an audio segment whose RMS is lower than the first threshold T1 is determined as invalid audio and directly filtered out (fail); an audio segment whose RMS is higher than the second threshold T2 is determined as valid speech and directly passed (pass); an audio segment whose RMS is between the two thresholds is determined as a boundary sample (boundary) and transferred to the second stage for fine discrimination.

[0054] In this embodiment, the first threshold T1 and the second threshold T2 are determined by the statistical distribution of the labeled dataset:

[0055] ;

[0056] ;

[0057] Construct a positive sample set S + and a negative sample set S -The dataset is labeled, where positive samples are valid speech samples and negative samples are invalid audio samples. The high quantile of the RMS distribution of the negative sample set is taken as T1, ensuring that the energy of most invalid audio samples is below this threshold and can be directly filtered out. The low quantile of the RMS distribution of the positive sample set is taken as T2, ensuring that the energy of most valid speech samples is above this threshold and can be directly allowed. The region between the two thresholds is the overlapping and undetermined region of the positive and negative sample energy distributions; only boundary samples within this region need to enter the second-level processing.

[0058] Step S4: Multidimensional acoustic feature extraction of boundary samples

[0059] Receive the boundary samples handed over in step S3, extract three types of preset human voice-specific acoustic features, and output the corresponding feature value set:

[0060] The first acoustic feature (periodicity strength feature): A frame-by-frame periodic detection is performed on the audio segment, and the segment-level feature value is obtained by statistically analyzing the frame-level detection results. The physical basis of this feature is: human voice is produced by the periodic vibration of the vocal cords, has a clear fundamental frequency and rich harmonics, and periodic detection can obtain stable results; environmental noise, current interference, etc., do not have periodicity, the detection results are disordered, and the feature value is significantly lower.

[0061] The second acoustic characteristic (fundamental frequency band energy proportion characteristic): Perform spectral analysis on the audio segment to calculate the ratio of the energy in the frequency band where the fundamental frequency of the human voice is located to the total energy of the segment. The physical basis for this characteristic is that the energy of the human voice is highly concentrated in the fundamental frequency and its harmonics, and the effective human voice accounts for a high proportion in this frequency band; the energy of broadband environmental noise is dispersed, and the energy of equipment noise deviates from this frequency band, accounting for a low proportion.

[0062] The third acoustic feature (low-frequency energy proportion feature): calculates the ratio of the energy of the preset low-frequency band to the total energy of the segment. The physical basis of this feature is: human voice is rich in low-frequency harmonic components and has a high low-frequency proportion; high-frequency dominant noises such as equipment howling and electrical interference have a low low-frequency proportion, which can complement the second acoustic feature and specifically capture high-frequency noise types that are difficult to distinguish by the second acoustic feature.

[0063] Step S5: Calculation and Judgment of Second-Level Weighted Score

[0064] Receive the feature value set of the three types of acoustic features from step S4, perform segmented level mapping and weighted scoring, and output the final judgment result. The specific process is as follows:

[0065] Piecewise ranking mapping: Each feature value is input into a dedicated piecewise step-type scoring mapping function. The continuous feature values ​​are discretized into normalized confidence levels, with a value range of [0, 1]. The higher the value, the more the feature supports the determination of valid human voice. The mapping function g1 for the first acoustic feature is: high level for strong human voice feature, medium level for weak human voice feature, and 0 for no human voice feature. The mapping function g2 for the second acoustic feature is: high level for high frequency band proportion, medium level for medium proportion, and 0 for low proportion. The mapping function g3 for the third acoustic feature is: high level for frequency band proportion higher than the segmentation threshold, and 0 otherwise.

[0066] Weighted summation for overall score: The overall score is calculated by fusing the confidence levels of each feature using a weighted scoring model. Where n is the number of features (n=3 in this embodiment); x i Let be the eigenvalue of the i-th feature; The segmented level mapping function for the corresponding features; w i Let be the weight of the i-th feature.

[0067] The weight of each feature is determined by its ability to distinguish between positive and negative samples; the stronger the ability to distinguish, the higher the weight. In this embodiment, the first acoustic feature has the strongest ability to distinguish and the highest weight, followed by the second acoustic feature, and the third acoustic feature has the lowest weight.

[0068] Threshold determination: The overall score is compared with the preset scoring threshold S. th Compare and output the judgment result:

[0069] ;

[0070] Step S6: Summarize and output the results

[0071] The results of the judgment of all audio segments in step S5 are summarized. Audio segments judged as valid speech are output to the downstream speech recognition engine, and audio segments judged as invalid audio are discarded. At the same time, the judgment label (pass / fail) and the comprehensive score of each segment are output for subsequent offline analysis and threshold optimization.

[0072] Taking an office scenario as an example, suppose three audio segments are obtained: segment A is valid speech from a user speaking directly into the microphone, segment B is air conditioner noise, and segment C is non-target speech from a colleague talking at a distance. After the first stage of energy pre-screening: the RMS of segment B is lower than T1 (weak air conditioner noise energy), so it is directly judged as invalid audio and discarded, with the output label "fail" and not further processed; the RMS of segment A is higher than T2 (sufficient energy from a user speaking at close range), so it is directly judged as valid speech and allowed to pass, with the output label "pass" and sent directly to the speech recognition engine; the RMS of segment C is between the two thresholds (medium energy of distant human voice, overlapping with some noise energy range), so it is transferred to the second stage as a boundary sample.

[0073] After the second-level feature extraction and scoring calculation, segment C had a moderate periodic feature score, but the energy proportion of the fundamental frequency band was low. The overall score did not reach the scoring threshold, so it was finally judged as invalid audio and filtered out, thus avoiding invalid recognition triggered by non-target speech.

[0074] Example 2:

[0075] This embodiment details the data-driven screening and determination process for the three key acoustic features in Embodiment 1. This process is reproducible and transferable, and the overall flow is as follows: Figure 2 As shown, the specific steps are as follows:

[0076] Step F1: Construct the labeled dataset

[0077] A diverse set of audio samples was collected from the target scene and manually labeled as positive and negative samples to construct a labeled dataset. Positive samples are valid speech whose semantics can be clearly discerned by the human ear, while negative samples are various invalid audio sources such as environmental noise, microphone interference, blurred audio, and device howling. There are hundreds to thousands of positive and negative samples in each category, covering typical acoustic conditions in the scene.

[0078] Step F2: Batch extraction of candidate acoustic features

[0079] For all samples in the labeled dataset, more than 100 candidate acoustic features are extracted in batches to form a fully covered candidate feature pool. The feature categories include: 1) Time domain features: zero crossing rate, short-time energy, autocorrelation coefficient, etc., about 10; 2) Frequency domain features: spectral centroid, spectral bandwidth, spectral roll-off point, etc., about 10; 3) Cepstral features: MFCC coefficients from order 1 to 13 and their statistics, about 40; 4) Energy features: RMS, energy proportion of each frequency band, etc., about 15; 5) Pitch features: fundamental frequency, maximum / minimum / average pitch, etc., about 10.

[0080] Step F3: Calculate the difference ratio of each candidate feature.

[0081] For each candidate feature j, calculate its difference ratio D between the positive and negative sample sets. j This is used to quantify the ability of this feature to distinguish between valid human voices and invalid audio. The calculation process is as follows:

[0082] Suppose that feature j has N values ​​in the positive sample set. + There are N observations on the negative sample set. - Each observation value. First, calculate the feature mean of the positive sample set. and the eigenmean of the negative sample set :

[0083] ;

[0084] ;

[0085] Then calculate the characteristic standard deviation of the positive sample set. and the characteristic standard deviation of the negative sample set :

[0086] ;

[0087] ;

[0088] The design philosophy of the discriminant ratio is that an excellent distinguishing feature should maximize the difference between the means of positive and negative samples while minimizing the dispersion of their respective distributions. Therefore, the discriminant ratio is calculated by dividing the difference in means by the joint standard error of the two sample mean estimates. for:

[0089] ;

[0090] In the formula, the numerator is the absolute difference between the means of positive and negative samples; the denominator is the joint standard error of the means estimates of the two samples, which characterizes the uncertainty of the means estimate. To avoid stability constants with a denominator of zero (take a very small positive number).

[0091] Difference ratio D j The physical meaning of D is: the ratio of the degree to which this feature separates the centers of positive and negative sample distributions to the degree of distribution overlap; j The larger the value, the less overlap there is between the two types of samples in that feature dimension, and the stronger the ability to distinguish them.

[0092] Step F4: Construct a feature importance ranking table

[0093] According to the difference ratio D j All candidate features are sorted in descending order from largest to smallest to construct a feature importance ranking table, which intuitively reflects the ranking of each feature's ability to distinguish between positive and negative samples.

[0094] Step F5: Select the optimal feature subset using multiple criteria

[0095] Based on the feature importance ranking table, the final optimal feature subset is obtained by jointly selecting according to the following four criteria: 1) High discriminative power criterion: setting a difference ratio threshold D. th Only keep D j >D thThe criteria include: 1) **Features that ensure the selected features have significant ability to distinguish between positive and negative samples;** 2) **Cross-scene stability criterion:** Calculate the difference ratio for candidate features under various typical noise scenarios, retain features with smaller variance in the difference ratio under each scenario, and remove unstable features that are only sensitive to specific noise and have poor generalization ability; 3) **Low redundancy criterion:** Calculate the Pearson correlation coefficient between each pair of features. ,when When a pair of features is identified as highly redundant, only the feature with the greater difference ratio is retained to avoid information overlap; 4) Computational efficiency criteria: Prioritize features that can be completed in a single traversal and have low time complexity, and eliminate features that require complex iterative calculations to ensure the real-time performance and edge deployment adaptability of the solution.

[0096] After the above screening, this embodiment ultimately yields three types of acoustic features, ranked by their difference ratios as follows: first acoustic feature (highest), second acoustic feature (second highest), and third acoustic feature (lowest). All three meet the four screening criteria, exhibiting orthogonal physical mechanisms, low information redundancy, and low computational cost. Furthermore, the difference ratio ranking directly determines the weight allocation in the weighted scoring model; features with higher difference ratios are assigned higher weights, achieving a quantitative match between discriminative ability and weight.

[0097] When the deployment scenario changes, you only need to collect the audio data of the new scenario and re-execute steps F1 to F5 to automatically export the feature combination and corresponding weights adapted to the new scenario. There is no need to redesign the filtering algorithm, which has strong portability.

[0098] Example 3:

[0099] This embodiment provides a two-stage cascaded speech quality filtering system that implements the above method, and its overall structure is as follows: Figure 3 As shown, the system is deployed at the front end of the speech recognition engine and includes a first-level filtering module 10, a second-level filtering module 20, and a result output module 30.

[0100] The first-level filtering module 10 receives the raw audio stream, performs preprocessing and speech activity segmentation, calculates the root mean square energy value of each audio segment, and performs dual-threshold three-class classification based on preset first and second thresholds. This module includes a VAD segmentation unit 11, an energy calculation unit 12, and an energy distribution unit 13. The VAD segmentation unit 11 receives the raw audio stream, performs sampling rate conversion and amplitude normalization preprocessing, detects speech activity intervals based on short-time energy and zero-crossing rate, segments the continuous audio stream into several independent audio segments, and outputs them to the energy calculation unit 12. The energy calculation unit 12 receives each audio segment, calculates the root mean square energy value of each segment, and outputs it to the energy distribution unit 13. The energy distribution unit 13 receives the root mean square energy value of each audio segment, compares it with preset first threshold T1 and second threshold T2, and performs three-class classification: audio segments determined to be invalid audio and their discard flags are output to the result output module 30; audio segments determined to be valid speech are directly output to the result output module 30; and the audio data of boundary samples are output to the second-level filtering module 20.

[0101] The second-level filtering module 20 receives boundary samples, extracts multi-dimensional human voice-specific acoustic features, and calculates a comprehensive score by weighting the data according to preset weights after segmentation and level mapping. The score is then compared with a preset scoring threshold, and the result is output. This module includes a feature extraction unit 21, a scoring calculation unit 22, and a threshold determination unit 23. The feature extraction unit 21 receives the audio data of the boundary samples and extracts three types of acoustic features according to the frequency band range corresponding to each acoustic feature: periodic intensity features, fundamental frequency band energy proportion features, and low-frequency energy proportion features. The corresponding feature values ​​are then output to the scoring calculation unit 22. The feature extraction unit 21 includes a pitch detection subunit and a spectrum analysis subunit, which respectively realize time-domain periodicity detection and frequency-domain energy analysis. The scoring calculation unit 22 receives the feature values ​​of three types of acoustic features, converts each feature value into a confidence level through a scoring mapping function with segmented thresholds corresponding to each feature, and then calculates the comprehensive score by weighting according to preset weights, and outputs it to the threshold determination unit 23. The weight of each feature is positively correlated with the difference ratio of the corresponding feature. The threshold determination unit 23 receives the comprehensive score, compares it with the preset scoring threshold, and outputs the determination result of valid speech or invalid audio to the result output module 30.

[0102] The result output module 30 is used to receive the audio segments and corresponding judgment results output by the first-level filtering module 10 and the second-level filtering module 20, summarize the final judgment of all audio segments, output the audio data judged as valid speech to the downstream speech recognition engine, and discard the audio judged as invalid; at the same time, it can output the judgment label and comprehensive score of each segment for subsequent analysis and optimization.

[0103] When the system is working, the original audio is input to the first-level filtering module 10. After segmentation, energy calculation and three-class classification, samples with clear energy are directly bypassed by the energy splitting unit 13 to the result output module 30 without entering the second level. Only boundary samples are sent to the second-level filtering module 20, where feature extraction, score calculation and threshold determination are completed in sequence before being fed into the result output module 30.

[0104] This system features full parameter adjustability: first threshold T1, second threshold T2, frequency band range of each acoustic feature, piecewise threshold of the scoring mapping function, preset weights, and scoring threshold S. th All parameters are configurable and can be adjusted according to the deployment scenario; the baseband detection algorithm can be selected from multiple algorithms such as autocorrelation method, YIN, and PYIN.

[0105] Application Results: Through real-world testing in various scenarios, including offices, dedicated microphones, and built-in computer microphones, this system effectively filters environmental noise, conversations among people, and low-quality audio, retaining only the clear voice of the target user. This results in a significant reduction in invalid recognition and a high retention rate of valid voice. Furthermore, the single-segment processing latency is at the millisecond level, significantly reducing the false trigger rate.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A two-stage cascaded speech quality filtering method based on multi-feature weighted scoring, characterized in that, Includes the following steps: S1. Perform sampling rate conversion and amplitude normalization preprocessing on the acquired raw audio stream to obtain a standardized audio signal; S2. For the standardized audio signal, based on short-time energy and zero-crossing rate to detect the speech activity interval, the continuous audio stream is divided into several independent audio segments; S3. Calculate the root mean square energy value for each audio segment, and perform a three-classification decision based on the root mean square energy value and a preset first threshold and second threshold: audio segments with root mean square energy values ​​lower than the first threshold are determined as invalid audio and are directly filtered; audio segments with root mean square energy values ​​higher than the second threshold are determined as valid speech and are directly passed; audio segments with root mean square energy values ​​between the first threshold and the second threshold are determined as boundary samples; wherein the first threshold is less than the second threshold. S4. Extract preset multidimensional human voice-specific acoustic features from the boundary samples to obtain the corresponding feature value set; S5. Perform segmented level mapping on each feature value in the feature value set, then sum the features according to the preset weights to obtain the comprehensive score, and compare the comprehensive score with the scoring threshold to obtain the final judgment result. S6. Summarize the judgment results of all audio segments, output the audio segments judged as valid speech to the downstream speech recognition engine, and discard the audio segments judged as invalid audio.

2. The two-stage cascaded speech quality filtering method based on multi-feature weighted scoring according to claim 1, characterized in that, The first threshold and the second threshold in step S3 are predetermined by the statistical distribution of the labeled dataset. Specifically, a labeled dataset is constructed that includes a set of valid positive speech samples and a set of invalid negative audio samples. The high quantile of the root mean square energy distribution of the negative sample set is taken as the first threshold, and the low quantile of the root mean square energy distribution of the positive sample set is taken as the second threshold. The interval between the two thresholds is the overlapping and undeterminable region of the positive and negative sample energy distributions.

3. The two-stage cascaded speech quality filtering method based on multi-feature weighted scoring according to claim 1, characterized in that, The multidimensional human voice-specific acoustic features in step S4 include a first acoustic feature, a second acoustic feature, and a third acoustic feature: The first acoustic feature is the periodic intensity of the audio segment, obtained by periodically detecting each frame of the audio segment and statistically analyzing the results. It is used to characterize the periodicity of the vocal cord vibration of the human voice. The second acoustic feature is the fundamental frequency band energy ratio, which is the ratio of the energy in the fundamental frequency band of the human voice in the audio segment to the total energy of the segment. It is used to characterize the concentration of human voice energy in the fundamental frequency band. The third acoustic feature is the low-frequency energy ratio, which is the ratio of the energy in the preset low-frequency band of the audio segment to the total energy of the segment. It is used to characterize the low-frequency energy distribution characteristics of the audio.

4. The two-stage cascaded speech quality filtering method based on multi-feature weighted scoring according to claim 3, characterized in that, Step S5 specifically includes: inputting the feature values ​​of each acoustic feature into the corresponding piecewise step-type scoring mapping function to obtain a confidence level with a value range of [0, 1]. The higher the confidence level value, the more the feature supports the determination that the current audio is a valid human voice; assigning a corresponding weight to each acoustic feature, the weight being positively correlated with the feature's ability to distinguish between positive and negative samples; and multiplying the confidence level of each acoustic feature by its corresponding weight and summing the results to obtain the comprehensive score of the audio segment.

5. The two-stage cascaded speech quality filtering method based on multi-feature weighted scoring according to claim 1, characterized in that, The multidimensional human voice-specific acoustic features in step S4 are pre-determined using a data-driven screening method. The screening process includes: collecting diverse audio samples from the target scene, manually labeling them as positive and negative samples to construct a labeled dataset; extracting five categories of candidate acoustic features—time domain, frequency domain, cepstral, energy, and pitch—from all samples in the labeled dataset to form a candidate feature pool; calculating the difference ratio of each candidate feature between the positive and negative sample sets, where the difference ratio is the ratio of the absolute difference between the mean values ​​of the positive and negative samples to the joint standard error of the estimated mean values ​​of the two samples, used to quantify the feature's ability to distinguish between valid human voices and invalid audio; arranging all candidate features in descending order of difference ratio to construct a feature importance ranking table; and selecting the optimal feature subset from the feature importance ranking table based on four screening criteria: high discriminative power, cross-scene stability, low redundancy, and computational efficiency, as the multidimensional human voice-specific acoustic features.

6. A two-stage cascaded speech quality filtering system based on multi-feature weighted scoring, characterized in that, It includes a first-level filtering module, a second-level filtering module, and a result output module; The first-level filtering module is used to receive the original audio stream, complete preprocessing and speech activity segmentation, calculate the root mean square energy value of each audio segment, and perform a dual-threshold three-class classification decision after passing the preset first threshold and second threshold. The audio segments judged as invalid audio and the discard flag are output to the result output module, the audio segments judged as valid speech are directly output to the result output module, and the audio segments judged as boundary samples are output to the second-level filtering module. The second-level filtering module is used to receive the boundary samples, extract multi-dimensional human voice-specific acoustic features, and calculate the comprehensive score by weighting the scores according to preset weights after segmentation and level mapping. After comparing with the scoring threshold, the judgment result and the corresponding audio segment are output to the result output module. The result output module is used to receive audio segments and corresponding judgment results output by the first-level filtering module and the second-level filtering module, and after summarizing, output the valid speech to the downstream speech recognition engine and discard the invalid audio.

7. The two-stage cascaded speech quality filtering system based on multi-feature weighted scoring according to claim 6, characterized in that, The first-level filtering module includes a VAD segmentation unit, an energy calculation unit, and an energy splitting unit. The VAD segmentation unit receives the original audio stream, performs sampling rate conversion and amplitude normalization preprocessing, detects speech activity intervals based on short-time energy and zero-crossing rate, segments the audio into several independent audio segments, and outputs them to the energy calculation unit. The energy calculation unit receives each audio segment, calculates the root mean square energy value of each audio segment, and outputs it to the energy splitting unit. The energy splitting unit is used to receive the root mean square energy value of each audio segment, compare it with the preset first threshold and second threshold, and then perform three-class splitting. Invalid audio and discard flags and valid audio segments are output to the result output module, and boundary samples are output to the second-level filtering module.

8. The two-stage cascaded speech quality filtering system based on multi-feature weighted scoring according to claim 6, characterized in that, The second-level filtering module includes a feature extraction unit, a scoring calculation unit, and a threshold determination unit. The feature extraction unit receives boundary samples and extracts three types of acoustic features according to the frequency band range corresponding to each acoustic feature: periodic strength features, fundamental frequency band energy proportion features, and low-frequency energy proportion features. The corresponding feature values ​​are then output to the scoring calculation unit. The scoring calculation unit receives the feature values ​​of the three types of acoustic features, performs segmented level mapping on each feature value using a scoring mapping function with segmented thresholds, and then sums the values ​​according to preset weights to obtain a comprehensive score, which is then output to the threshold determination unit. The threshold determination unit is used to receive the comprehensive score, compare it with the scoring threshold, and output the result of determining whether the corresponding audio segment is valid or invalid audio to the result output module.

9. The two-stage cascaded speech quality filtering system based on multi-feature weighted scoring according to claim 8, characterized in that, The preset weights of each acoustic feature in the scoring calculation unit are positively correlated with the difference ratio of the corresponding feature; the difference ratio is the ratio of the absolute difference of the mean of the feature on the positive sample set and the negative sample set to the joint standard error.

10. The two-stage cascaded speech quality filtering system based on multi-feature weighted scoring according to claim 8, characterized in that, The system is deployed at the front end of the speech recognition engine; the first threshold, the second threshold, the frequency band range of each acoustic feature, the segmented threshold of the scoring mapping function, the preset weight, and the scoring threshold are all configurable parameters that can be adjusted according to the deployment scenario.

Images