Speech recognition method for expense reimbursement speech entry

Abstract

The present application relates to the technical field of speech signal processing, and discloses a speech recognition method for expense reimbursement speech input, comprising: generating a digitized audio stream from a continuous speech signal buffer; constructing a reference feature frame stream by extracting a mel-frequency cepstral coefficient feature sequence through a short-time Fourier transform and storing the reference feature frame stream in a storage array; calculating an energy difference value and a spectral flux variability of adjacent reference feature frames, marking a candidate phoneme anchor frame when the energy difference value and the spectral flux variability exceed a threshold value, and writing a timestamp into a control register; adjusting a feature sampling step size according to a timestamp difference value of adjacent candidate phoneme anchor frames; splicing and constructing a feature tensor according to an original time sequence from feature frames extracted at different step sizes, and inputting the feature tensor into an acoustic decoding model to output a reimbursement text; and the present application uses double indicators to determine phoneme mutation boundaries and dynamically adjusts a sampling step size, blocks noise aliasing to eliminate phoneme swallowing, and simultaneously shrinks a feature tensor size to reduce a calculation overhead and an inference delay.

Description

Technical Field

[0001] This invention relates to the field of speech signal processing technology, and more specifically, to a speech recognition method for speech input for expense reimbursement. Background Technology

[0002] Current continuous speech signal processing technology typically employs a sliding window mechanism with a fixed time step to extract continuous speech. It then extracts a feature frame stream of equal length and weighted Mel-Cepstral Coefficients through Short-Time Fourier Transform. The feature sequence, which is uniformly distributed throughout the time period, is input into the back-end acoustic decoding model to complete acoustic feature modeling. This strategy constitutes the conventional solution adopted by the current system. However, in the application of financial data voice input, continuous speech has non-uniform temporal distribution characteristics. Key entity acoustic segments such as monetary values, product names, and units of measurement are sparse on the time axis and experience drastic instantaneous energy fluctuations. Meanwhile, interjections, environmental noise, and hesitations occupy a large amount of temporal space.

[0003] Existing technologies use a fixed sliding window step size to extract features of equal length. This leads to the sparse, high-value phoneme features being aliased and diluted by non-critical information in the feature stream. This causes the decoder to confuse, replace, and miss key digital entities when faced with continuous long sentences. Reducing the sliding window step size to increase sampling density results in an increase in the size of the feature sequence, causing hardware processing overhead to exceed the carrying capacity of mobile devices. Furthermore, backend error correction cannot reverse the distortion of the frontend signal. For example, Chinese invention patent application CN120783732A discloses an intelligent speech recognition system and method based on the Internet of Things. It obtains the dynamic recognition step size per unit word segment by calculating the ratio of sentence recognition step size to the total number of word segments, and uses the time difference to construct feedback parameters to implement online recognition correction. This control method implicitly relies on a relatively clean acoustic environment and the speech sequence being accurately pre-defined. The ideal premise of word segmentation is that, in actual expense reimbursement voice input scenarios, key entities such as monetary values ​​and product names are extremely sparse in the time domain and experience violent transient energy fluctuations, accompanied by high-frequency bursts of non-speech parasitic noise such as keyboard tapping and paper tearing. In this non-ideal, noisy environment, the overall semantic word segmentation and template matching mechanism suffers a fundamental mismatch due to the nonlinear distortion of the front-end signal itself. It cannot implement sensitive step size adjustment at the boundary of surface phoneme mutations and is prone to oversampling oscillations caused by noise pseudo-energy, leading to phoneme swallowing and decoding error propagation. Since existing improvement strategies are limited by uniform feature extraction architecture and single signal adjustment boundary, a non-uniform temporal acoustic feature tensor extraction mechanism with adaptive adjustment characteristics is constructed in a noisy environment with discontinuous human-computer interaction characteristics to eliminate the entropy redundancy of the high-dimensional feature space.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a dual physical index parallel criterion by using the absolute value of the short-time logarithmic energy difference between adjacent speech frames and the frequency domain spectral flux variability, adaptively reconstruct the non-uniform temporal acoustic feature tensor, and establish a two-layer steady-state loop of feedforward hindrance and feedback gain to achieve efficient phoneme alignment under noisy conditions. Summary of the Invention

[0005] To address the problems in the background art, this invention provides a speech recognition method for voice input of expense reimbursement, comprising the following steps: Step S101: The acquired continuous speech signal is buffered into a data buffer to generate a digital audio stream; Step S102: Perform short-time Fourier transform on the digitized audio stream, extract the multidimensional Mel-Cepstral coefficient feature sequence, construct the reference acoustic feature frame stream and store it in the storage array; Step S103: Read adjacent reference acoustic feature frames, calculate the acoustic energy difference and frequency domain spectral flux variability between adjacent reference acoustic feature frames; and when the acoustic energy difference exceeds a preset energy threshold or the frequency domain spectral flux variability exceeds a preset variability threshold, mark the corresponding reference acoustic feature frame as a candidate entity phoneme anchoring frame on the time axis and write its timestamp into the control register. Step S104: Read the timestamp difference between adjacent candidate entity phoneme anchoring frames in the control register, and adjust the feature sampling step size according to the geometric distribution density reflected by the timestamp difference: In the high-density target interval where the temporal interval between two adjacent candidate entity phoneme anchoring frames is less than the preset time window, reduce the feature sampling step size to the first step size; In the low-density redundant interval where there are no unmarked candidate entity phoneme anchoring frames, expand the feature sampling step size to the second step size, where the first step size is smaller than the second step size. Step S105 involves concatenating the acoustic feature frames extracted by different feature sampling steps in their original temporal sequence to construct a discrete temporal acoustic feature tensor, which is then input into the acoustic decoding model to complete the alignment decoding and output the reimbursement text data.

[0006] Preferably, while extracting the multidimensional Mel-Cepstral coefficient feature sequence by performing short-time Fourier transform on the digitized audio stream, the transient signal-to-noise ratio (SNR) of the digitized audio stream in the full frequency domain is monitored in real time. When the transient SNR drops by more than 15 dB within 10 ms, it is determined that there is a sudden non-speech parasitic noise in the digitized audio stream. The marking action of anchoring the candidate entity phoneme is then suspended, and the feature sampling step size is controlled to remain unchanged at the current value to block oversampling caused by noise pseudo-energy, until the transient SNR in the full frequency domain recovers to above the safe threshold.

[0007] Preferably, in step S103, when calculating the acoustic energy difference value, parameter D is configured as the absolute value of the difference between the short-time logarithmic energy value of the reference acoustic feature frame at the current position and the short-time logarithmic energy value of the reference acoustic feature frame at the previous adjacent position; when the acoustic energy difference value exceeds the preset energy threshold, the reference acoustic feature frame at the corresponding position is marked as a candidate entity phoneme anchoring frame.

[0008] Preferably, step S104 further includes the following sub-steps: step S1041, using a counter register to accumulate the duration of continuous use of the second step; step S1042, when the duration exceeds 300ms, controlling the feature sampling step size to be downsampled from the second step size back to a fixed value of 15ms until the next candidate entity phoneme anchor frame that meets the threshold condition is detected.

[0009] Preferably, the preset time window is 50ms, the first step is 5ms long, and the second step is 30ms long; in the low-density interval of the unlabeled candidate entity phoneme anchor frame, sparse sampling is implemented through the second step to discard the background noise of the environment.

[0010] Preferably, step S105 further includes the following sub-steps: step S1051, obtaining the first acoustic feature frame sequence extracted using the first step length within the high-density target interval, and obtaining the second acoustic feature frame sequence extracted using the second step length within the low-density redundant interval; step S1052, extracting the original timestamps corresponding to each first acoustic feature frame sequence and each second acoustic feature frame sequence; step S1053, concatenating and splicing the first acoustic feature frame sequence and the second acoustic feature frame sequence according to the original timestamps from early to late to construct a discrete temporal acoustic feature tensor, so as to reduce the size of the feature matrix input to the acoustic decoding model.

[0011] Preferably, the calculation of frequency domain spectral flux variability in step S103 further includes the following sub-steps: Step S1031, obtaining the frequency domain spectral distribution vector of the reference acoustic feature frame at the current position on the time axis, and obtaining the frequency domain spectral distribution vector of the reference acoustic feature frame at the previous adjacent position; Step S1032, calculating the energy change of the frequency domain spectral distribution vector at the current position and the frequency domain spectral distribution vector at the previous adjacent position in each frequency domain channel; Step S1033, summing the squares of the energy changes in each frequency domain channel to generate frequency domain spectral flux variability.

[0012] Preferably, the storage array is a dual-port static random access memory array; a reference acoustic feature frame stream is cyclically written to the storage array through the first port, and adjacent reference acoustic feature frames are periodically read from the storage array through the second port.

[0013] Preferably, step S104, which adjusts the feature sampling step size based on the timestamp difference between adjacent candidate entity phoneme anchor frames in the control register, includes: when the timestamp difference between two adjacent candidate entity phoneme anchor frames is less than a preset time window, the current feature sampling step size is switched to the first step size.

[0014] Preferably, in step S105, the acoustic decoding model includes a deep neural network architecture and a connectionist temporal classification decoder; the discrete temporal acoustic feature tensor is derived using the acoustic decoding model, and the corresponding reimbursement text data is output.

[0015] The embodiments of the present invention have at least the following beneficial effects: 1. In the speech recognition of expense reimbursement voice input, the temporal analysis module establishes a dual physical index parallel criterion by calculating the absolute value of the short-time logarithmic energy difference and the spectral flux variability between adjacent speech frames, and calibrates the abrupt boundary of specific phonemes on the time axis. Based on the local geometric distribution density of the abrupt boundary on the time axis, the window controller adaptively reduces the feature sampling step size to 5ms in the high-density interval with a time interval of less than 50ms to implement local oversampling. The control mechanism blocks the high-value phoneme features from being mixed with adjacent background noise in the feature stream throughout the time period, ensuring that the acoustic decoding model obtains phoneme detail features with clear boundaries, and eliminating phoneme swallowing and misspelling faults caused by non-uniform temporal distribution.

[0016] 2. To address the feature space redundancy caused by non-critical background acoustic information, the window controller adaptively expands the feature sampling step size to 30ms in the low-density redundant interval of the unlabeled candidate entity phoneme anchor frame to implement sparsity downsampling, discarding ambient background noise and tone reverberation. The feature splicing unit concatenates the feature frames extracted by different step sizes in the original temporal order to construct a non-uniform discrete temporal acoustic feature tensor with global dimensional shrinkage. The data stream undergoes structural shrinkage in the temporal domain, which significantly reduces the tensor size input to the acoustic decoding model and greatly reduces the matrix operation overhead, thereby reducing the decoding computational entropy and inference latency.

[0017] 3. To address the pseudo-energy mutations caused by high-frequency sudden mechanical noise in practical applications, the pre-filter simultaneously monitors the transient signal-to-noise ratio (SNR) across the entire frequency domain in real time while extracting the multidimensional Mel-frequency cepstral coefficient feature sequence. When the transient SNR drops by more than 15 dB within 10 ms, it is determined that the system has encountered sudden parasitic interference. The timing analysis module then suspends the marking action of anchoring candidate entity phonemes, and the window controller forcibly maintains the current feature sampling step size constant until the transient SNR recovers to a safe range. This cuts off the computational resource consumption caused by unnecessary oversampling. Based on the adaptive adjustment mechanism of the feature reconstruction process, the stability of the system extraction state is maintained. Attached Figure Description

[0018] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the invention are illustrated by way of example and not limitation, wherein: Figure 1 This is a flowchart of the expense reimbursement voice recognition method based on dynamic feature sampling according to the present invention; Figure 2 This is a multi-level functional architecture diagram of the expense reimbursement voice recognition system of the present invention. Detailed Implementation

[0019] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments in conjunction with the accompanying drawings. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0020] A speech recognition method for voice input in expense reimbursement includes the following steps: Step S101: The acquired continuous speech signal is buffered into a data buffer to generate a digital audio stream; Step S102: Perform short-time Fourier transform on the digitized audio stream, extract the multidimensional Mel-Cepstral coefficient feature sequence, construct the reference acoustic feature frame stream and store it in the storage array; Step S103: Read adjacent reference acoustic feature frames, calculate the acoustic energy difference and frequency domain spectral flux variability between adjacent reference acoustic feature frames; and when the acoustic energy difference exceeds a preset energy threshold or the frequency domain spectral flux variability exceeds a preset variability threshold, mark the corresponding reference acoustic feature frame as a candidate entity phoneme anchoring frame on the time axis and write its timestamp into the control register. Step S104: Read the timestamp difference between adjacent candidate entity phoneme anchoring frames in the control register, and adjust the feature sampling step size according to the geometric distribution density reflected by the timestamp difference: In the high-density target interval where the temporal interval between two adjacent candidate entity phoneme anchoring frames is less than the preset time window, reduce the feature sampling step size to the first step size; In the low-density redundant interval where there are no unmarked candidate entity phoneme anchoring frames, expand the feature sampling step size to the second step size, where the first step size is smaller than the second step size. Step S105 involves concatenating the acoustic feature frames extracted by different feature sampling steps in their original temporal sequence to construct a discrete temporal acoustic feature tensor, which is then input into the acoustic decoding model to complete the alignment decoding and output the reimbursement text data.

[0021] Preferably, while extracting the multidimensional Mel-Cepstral coefficient feature sequence by performing short-time Fourier transform on the digitized audio stream, the transient signal-to-noise ratio (SNR) of the digitized audio stream in the full frequency domain is monitored in real time. When the transient SNR drops by more than 15 dB within 10 ms, it is determined that there is a sudden non-speech parasitic noise in the digitized audio stream. The marking action of anchoring the candidate entity phoneme is then suspended, and the feature sampling step size is controlled to remain unchanged at the current value to block oversampling caused by noise pseudo-energy, until the transient SNR in the full frequency domain recovers to above the safe threshold.

[0022] Preferably, in step S103, when calculating the acoustic energy difference value, parameter D is configured as the absolute value of the difference between the short-time logarithmic energy value of the reference acoustic feature frame at the current position and the short-time logarithmic energy value of the reference acoustic feature frame at the previous adjacent position; when the acoustic energy difference value exceeds the preset energy threshold, the reference acoustic feature frame at the corresponding position is marked as a candidate entity phoneme anchoring frame.

[0023] Preferably, step S104 further includes the following sub-steps: step S1041, using a counter register to accumulate the duration of continuous use of the second step; step S1042, when the duration exceeds 300ms, controlling the feature sampling step size to be downsampled from the second step size back to a fixed value of 15ms until the next candidate entity phoneme anchor frame that meets the threshold condition is detected.

[0024] Preferably, the preset time window is 50ms, the first step is 5ms long, and the second step is 30ms long; in the low-density interval of the unlabeled candidate entity phoneme anchor frame, sparse sampling is implemented through the second step to discard the background noise of the environment.

[0025] Preferably, step S105 further includes the following sub-steps: step S1051, obtaining the first acoustic feature frame sequence extracted using the first step length within the high-density target interval, and obtaining the second acoustic feature frame sequence extracted using the second step length within the low-density redundant interval; step S1052, extracting the original timestamps corresponding to each first acoustic feature frame sequence and each second acoustic feature frame sequence; step S1053, concatenating and splicing the first acoustic feature frame sequence and the second acoustic feature frame sequence according to the original timestamps from early to late to construct a discrete temporal acoustic feature tensor, so as to reduce the size of the feature matrix input to the acoustic decoding model.

[0026] Preferably, the calculation of frequency domain spectral flux variability in step S103 further includes the following sub-steps: Step S1031, obtaining the frequency domain spectral distribution vector of the reference acoustic feature frame at the current position on the time axis, and obtaining the frequency domain spectral distribution vector of the reference acoustic feature frame at the previous adjacent position; Step S1032, calculating the energy change of the frequency domain spectral distribution vector at the current position and the frequency domain spectral distribution vector at the previous adjacent position in each frequency domain channel; Step S1033, summing the squares of the energy changes in each frequency domain channel to generate frequency domain spectral flux variability.

[0027] Preferably, the storage array is a dual-port static random access memory array; a reference acoustic feature frame stream is cyclically written to the storage array through the first port, and adjacent reference acoustic feature frames are periodically read from the storage array through the second port.

[0028] Preferably, step S104, which adjusts the feature sampling step size based on the timestamp difference between adjacent candidate entity phoneme anchor frames in the control register, includes: when the timestamp difference between two adjacent candidate entity phoneme anchor frames is less than a preset time window, the current feature sampling step size is switched to the first step size.

[0029] Preferably, in step S105, the acoustic decoding model includes a deep neural network architecture and a connectionist temporal classification decoder; the discrete temporal acoustic feature tensor is derived using the acoustic decoding model, and the corresponding reimbursement text data is output.

[0030] Example 1: The current expense reimbursement voice input system is configured in a financial office environment. The system includes an audio acquisition module, a Mel-Cepstral Coefficient (MCC) processing unit, a time-series analysis module, a window controller, and an acoustic decoding model. The audio acquisition module acquires continuous speech signals, converts them into a digital audio stream, and stores them in a time-domain data buffer. The MCC processing unit performs a short-time Fourier transform on the digital audio stream, extracts the MCC feature sequence, constructs a reference acoustic feature frame stream, and stores it in a storage array. The time-series analysis module periodically reads adjacent reference acoustic feature frames and calculates the acoustic energy difference between the current frame and the previous frame. and spectral flux variability The spectral flux variability Within the data logic domain, this is defined as the energy change scalar on the frequency domain distribution vector between two adjacent temporal acoustic feature frames. Specifically, the temporal analysis module obtains the frequency domain spectral distribution vector composed of the 512-dimensional logarithmic power spectral density components of the current frame on the time axis, and obtains the vector of the same dimension from the previous adjacent frame. In each independent frequency channel, the temporal analysis module subtracts the logarithmic power spectral scalar of the previous frame from the current frame's logarithmic power spectral scalar to calculate the energy change in each channel. Subsequently, the processing logic squares the energy changes in all 512 channels and sums the squared values, ultimately outputting a positive scalar representing the nonlinear fluctuation of the spectral structure envelope, thus characterizing the transient boundary spectral flux variability of phonemes in the frequency domain. The calculation formula is as follows: , ,in, This is the short-time logarithmic energy value of the current frame. This represents the short-time logarithmic energy value of the preceding frame. This is the frequency domain spectral distribution vector of the current frame. This is the frequency domain spectral distribution vector of the preceding frame. For the L2 distance, when or At that time, the reference acoustic feature frame is marked as the candidate entity phoneme anchor frame, and the corresponding timestamp is written into the control register, where To preset the energy threshold, This is a preset threshold for spectral flux variability.

[0031] The window controller reads the timestamp difference from the control register and adjusts the feature sampling step size. Specifically, the geometric distribution density is calculated as follows: a sliding observation time window of 50 milliseconds is set, and the total number of consecutive candidate entity phoneme anchoring frames within the observation time window is calculated. Then, the total number is divided by the duration of the observation time window to obtain the phoneme mutation frequency scalar per unit time. When the frequency scalar is greater than or equal to 40 per second, it is determined that the phoneme distribution within the time period exhibits a high geometric distribution density. The window controller then triggers sampling step size contraction control. When the time interval between two adjacent candidate entity phoneme anchoring frames is less than 50ms, it is determined that the system is in a high-density data entry state, and the feature sampling step size is set to 5ms to perform oversampling to preserve data. To preserve the transient details of financial entity phonemes, within the low-density redundant interval of the unlabeled candidate entity phoneme anchor frames, the window controller sets the feature sampling step size to 30ms, performing sparse downsampling to suppress environmental noise and vocal reverberation. The feature concatenation unit concatenates the feature frames extracted at the above step size according to the original time sequence to construct a non-uniform discrete temporal acoustic feature tensor, which is input into the acoustic decoding model for character-aligned decoding, outputting reimbursement text data. To maintain the alignment convergence of the decoding model under non-uniform temporal distribution, this invention configures a time-step reassembly network layer after the input layer of the acoustic decoding model. The network layer reads the timestamp metadata in the control register and identifies the boundary time interval between the high-density sampling region and the low-density sampling region in the current input feature frame sequence. Next, for high-density feature frames with a sampling step of 5 milliseconds, local downsampling to a 15-millisecond reference time value is performed using a one-dimensional linear convolution operator. For low-density feature frames with a sampling step of 30 milliseconds, local upsampling is performed using a bilinear interpolation operator between adjacent frames in the time domain. This normalizes the interwoven high- and low-density feature tensors into a standard time-series matrix with a uniform step of 15 milliseconds at the data structure level, ensuring that the deep neural network architecture can perform continuous matrix multiplication and state propagation according to causal relationships. The time-step reorganization network layer establishes a time-domain normalization architecture for discrete feature frames based on the signal sampling rate transformation theory and the constraint of uniform resampling of discrete-time signals. The input receives the discrete-time acoustic feature tensor and the timestamp data stored in the control register. The data, discrete temporal acoustic feature tensors, are composed of multidimensional Mel-Cepstral coefficient feature sequences. The processing stage includes reading timestamp metadata, identifying abrupt change nodes interwoven with different feature sampling steps on the time axis, dividing high-density sampling regions and low-density sampling regions. For the high-density sampling region with a feature sampling step of 5 milliseconds, a one-dimensional linear convolution operator slides along the time axis with a window step of 15 milliseconds to compress the temporal resolution. For the low-density sampling region with a feature sampling step of 30 milliseconds, a bilinear interpolation operator extracts the temporal weighted average of adjacent acoustic feature frames to fill in the temporal gaps. By reducing the scale of the original data stream of non-critical audio segments through sparse sampling, the non-critical background speech data in the recording stream is contracted within the acquisition boundary, thereby achieving local privacy protection of the audio data stream.The output generates a 15-millisecond standard time-series matrix that is uniformly distributed in the time domain, which is directly fed into the deep neural network architecture.

[0032] To address transient parasitic interference, the pre-filter monitors the signal-to-noise ratio (SNR) of the digitized audio stream in real time. When the SNR drops by more than 15 dB within 10 ms, the timing analysis module stops marking candidate entity phoneme anchor frames. The window controller maintains the current feature sampling step size until the SNR recovers. For scenarios with trailing or hesitant sounds, a counter register accumulates the duration of the 30 ms sampling step. When the duration exceeds 300 ms, the window controller adjusts the sampling step size to 15 ms for uniform sampling until the candidate entity phoneme anchor frame is recaptured, thus avoiding critical audio interference. In the event of element loss, the window controller, based on the hysteresis circulation mechanism and steady-state switching criterion of control theory, establishes a step size adjustment latch control loop to cope with threshold critical point signal disturbances. The input end collects the timestamp difference between the current feature sampling step size and the anchor frames of two adjacent candidate entity phonemes in real time. The processing stage includes: when the timestamp difference triggers the feature sampling step size to switch between 5 milliseconds and 30 milliseconds, the time counter starts to accumulate, limiting the sampling step size switching frequency. The minimum duration after switching is set to 200 milliseconds, and the acoustic energy difference value is stopped within the 200-millisecond time span. With frequency domain spectral flux variability The threshold comparison maintains a constant feature sampling step size; the output terminal outputs the feature sampling step size, and the threshold judgment control authority is restored after the time counter exceeds 200 milliseconds, eliminating the non-steady-state response of the sampling stream oscillating repeatedly at the critical point. In practical engineering applications, the transient time window for judging sudden parasitic noise is set to 10 milliseconds and the drop amplitude threshold is set to 15 dB. The basis for this is that when real speech streams in human office environments emit plosive sounds or pause and switch, the maximum energy attenuation slope of their full-frequency domain signal-to-noise ratio within 10 milliseconds usually does not exceed 8 dB. However, non-speech transient parasitic interferences such as violent keyboard typing and paper tearing have pulse-type high-energy burst characteristics, and their energy will produce an extremely steep step drop within 10 milliseconds, with the drop amplitude generally exceeding 15 dB. Therefore, the group quantization threshold can accurately identify non-speech noise and block unnecessary oversampling.

[0033] Example 2: The current experiment was conducted in an acoustically isolated office environment. Speech signals were acquired using an omni-directional test microphone array to construct a test set containing the original expense reimbursement data. The ambient noise floor power density was set to -60dBm, and Gaussian white noise with a signal-to-noise ratio of 18.4dB was simultaneously superimposed on the input path to simulate typical interference in an open office space. The audio acquisition module converted the speech signal into a digital audio stream at a sampling frequency of 16kHz and sent it to the time-series analysis module for real-time feature processing. To demonstrate the effectiveness of the proposed solution under complex expense reimbursement conditions, four sample groups were set up: the experimental group used a dynamic sampling step size adjustment mechanism; control group one used a sliding window with a fixed sampling step size of 10ms; control group two limited the feature sampling step size to 35ms to observe the smoothing effect of excessively long windows on feature details; control group three set the energy threshold... The value is set to 0.12, which exceeds the limits of the present invention. The upper limit of the range was used to observe the impact of high-sensitivity triggering on system load fluctuations. Each sample group continuously input voice samples containing the reimbursement of 152.5 yuan for transportation expenses. The processing terminal synchronously recorded the end-to-end inference delay and recognition error rate from audio input to output text. In this invention, the time-series analysis module periodically reads adjacent reference acoustic feature frames and calculates the energy difference value. With spectral flux variability The calculation formula is as follows: , ,in, This is the short-time logarithmic energy value of the current frame. This represents the short-time logarithmic energy value of the preceding frame. This is the current frame frequency domain spectral distribution vector. The frequency domain spectral distribution vector of the preceding frame. For the L2 distance, when or At that time, the system marks the candidate entity phoneme anchor frame and writes the timestamp to the control register, where To preset the energy threshold, This is a preset threshold for spectral flux variability.

[0034] The test results show that when the financial keyword "transportation expenses" is detected, When the instantaneous value reaches 0.45, the anchoring determination is triggered. Based on the timestamp information in the control register, the window controller adaptively switches the feature sampling step size from 30ms to 5ms. Under the effect of this invention, the feature sampling density of the keyword region reaches 200 frames / second, compared to 100 frames / second in control group one, reducing feature space redundancy by 42.1%. Control group three... The setting was too high, and the anchor frame was missed due to energy fluctuations caused by slight accents. The sampling step size was not switched in time, resulting in the recognition error of the digits 1 and 5 in 152.5, with an error rate of 8.5%. The error rate of the present invention was 1.2%. The control group 2 used a fixed sampling step size of 35ms, which caused feature blurring in high-frequency speech segments, with an average recognition delay of 280ms. The average recognition delay of the present invention was 45.6ms.

[0035] The above data shows that the control strategy based on the dynamic switching of sampling step size of the candidate entity phoneme anchor frame, by performing local high-density oversampling on key phonemes, suppresses environmental background noise interference while preserving transient details. The data trend of control group three shows that... When the range deviates from the upper limit of the present invention, the feature acquisition is delayed due to the lack of anchoring, which verifies the necessity of the numerical range defined by the present invention to ensure the stability of acoustic decoding. The present invention achieves the optimal match between the acoustic feature tensor size and the decoding accuracy through nonlinear control of feature sampling density.

[0036] Example 3: The current expense reimbursement voice input system is configured on an office terminal. The system includes a digital signal processing unit, a dynamic window adjuster, and a character decoding model. The digital signal processing unit preprocesses the continuous voice input into a digital audio stream and constructs a time-series feature sequence through Mel-frequency cepstral transform. The sequence is stored in a memory buffer. The system has a built-in real-time operating condition monitoring module that monitors the energy distribution of the input audio stream in the time domain and sets a sensitivity threshold for energy changes. The value is set to 0.06 to identify abrupt boundaries in speech segments. When the energy difference between adjacent speech frames is greater than 0.06, the threshold value is determined. At that time, the logic unit marks the frame as the starting point of the entity phoneme feature and records its timestamp to update the feature sampling step size.

[0037] To address the data redundancy issue in speech recognition, the dynamic window adjuster is configured with sampling step size switching logic. When the transient feature region of an entity phoneme is detected, the system adaptively reduces the sampling step size to 5ms to capture the instantaneous details of high-frequency phonemes such as numbers and units in financial reimbursement scenarios. This process achieves data smoothing by reducing feature extraction latency. When the steady-state region dominated by filler words or environmental noise is detected, the system adaptively expands the sampling step size to 30ms. This suppresses irrelevant acoustic disturbances through sparse sampling. The feature splicing unit reads the non-uniform sampling step size feature frames from the cache and constructs a feature with dimension 1. The discrete-time acoustic feature tensor, where... For the feature dimension of the Mel spectrum, Given the total number of feature frames, a tensor is input to the character decoding model. The model calculates the weights of feature components through an attention mechanism and outputs text data. In actual system operation, the model adopts a dual-branch hybrid decoding architecture that combines temporal classification and attention mechanisms. The connectionist temporal classification decoder serves as the main decoding branch, performing explicit monotonic temporal frame alignment operations on the input discrete temporal acoustic feature tensor and outputting coarse-grained text candidate sequences. At the same time, the attention mechanism serves as an auxiliary calibration branch, calculating the global weights of feature components in the temporal domain. This is used to perform long-distance semantic reclassification and contextual error correction on the candidate sequences output by the temporal classification decoder. The two are deduced collaboratively through a weighted loss function, thereby improving the accuracy of the final text by utilizing the attention mechanism without violating the monotonic alignment rule.

[0038] To smooth out time-domain distortion caused by sampling step size switching, the system introduces a phase consistency compensation operator. : ,in, The feature vector components within 5 frames before and after the sampling step size switching point are used. The number of frames participating in the compensation and , The compensation operator is the arithmetic mean of the eigenvector components within the window. The feature matrix of the splicing nodes is calibrated to counteract spectral aliasing caused by abrupt changes in step size. The logic control unit verifies the semantic confidence of the output text in real time. If the confidence is below 0.85, the window controller forces the sampling step size back to 2ms and increases the fine-grained feature acquisition frequency until the confidence recovers to above 0.90. The values ​​of the above confidence threshold and backtracking step size were determined in a financial environment simulation test by balancing decoding latency and word error rate. Engineering tests show that when the text semantic confidence drops below 0.85, the table... The non-uniform sampling of 5ms or 30ms has already caused the omission of high-frequency digital phonemes. At this time, forcibly reducing the sampling step size to the extreme value of 2ms allows the system to rescan the audio stream at an ultra-high density of 500 frames per second in the time domain, forcibly capturing the transient weak acoustic edges of Chinese plosives and monetary units, thereby achieving oversaturation acquisition of fine-grained features and ensuring that the confidence level can quickly recover to the safe decoding range above 0.90. The system achieves acoustic feature alignment in high-noise environments through a two-layer loop of feedforward compensation and feedback calibration.

[0039] Example 4: The current expense reimbursement voice input system configuration includes a signal preprocessing unit, a feature converter, and an acoustic decoding model. Upon startup, the system performs an environmental noise baseline calibration process. The system remains silent while reading a 500ms raw audio stream and calculates the average noise power spectral density of each frequency band within the time period using a short-time Fourier transform. The logic control unit is based on Set the lower limit of signal-to-noise ratio tolerance The calculation formula is: ,in, As a preset benchmark mean for speech features, when the signal-to-noise ratio of the real-time audio input is lower than... At this time, the system automatically increases the sensitivity of energy detection to counteract the interference of background noise on the anchor frame marker of candidate entity phonemes. This calibration data is stored in a non-volatile register for subsequent reimbursement entry tasks to ensure the consistency of the system's state under different environmental conditions.

[0040] To address transient interference such as keystrokes and paper tearing in reimbursement scenarios, the feature converter incorporates a negative event feature mapping table. This table stores discrete feature vectors of the interfering acoustic signals, and the system calculates the cosine similarity between the current audio frame and the feature vectors in the mapping table in real time. The calculation formula is: ,in, This is the cepstral feature vector of the current audio frame. The interference feature vector in the mapping table, For the L2 distance, when At that time, the logic control unit determines the current frame as a negative event and generates an interference timestamp and interference intensity coefficient. Inject decoding path, where The acoustic decoding model uses the Euclidean distance between the current feature vector and the negative event feature vector in the interference trajectory buffer as a reference. The system records and weights the affected feature components to ensure that the alignment decoding process only operates on the pure speech phoneme tensor, thereby achieving implicit defense against non-speech noise. Specifically, the interference intensity coefficient, calculated by the Euclidean distance between two vectors, represents the severity of interference in the current frame. To inject this scalar coefficient into the high-dimensional feature space, this invention uses a linear mapping weight matrix from 1 to 512 dimensions to expand the interference intensity coefficient into a gain control vector that is completely consistent with the dimension of the Mel spectrum feature. The gain control vector is then multiplied element-wise with the feature vector of the current audio frame, and the energy of the interfered channel is selectively attenuated through a subtraction mechanism, thereby achieving implicit defense against non-speech noise at the high-dimensional feature matrix level.

[0041] Example 5: The current expense reimbursement voice input system configuration includes a signal processing module, a feature tensor builder, a confidence evaluator, and a decoding network. To ensure consistency in the acoustic environment, the system is configured based on the standard ambient noise level before deployment. After establishing a reference baseline and initiating the offline environmental measurement process, the audio acquisition module continuously samples for 1000ms in a silent state to obtain the raw audio signal stream composed of environmental noise. The signal processing module converts the audio stream to the frequency domain using a fast Fourier transform and calculates the environmental reference spectrum distribution vector. The unit is dB, and it is stored in a non-volatile memory array. During system initialization, the input speech frame spectrum distribution vector is calculated. and Euclidean distance between As a verification criterion for determining whether the current environment has shifted, when Exceeding the offset threshold At this time, the system triggers a zero-point calibration procedure, using the current environmental noise distribution as a new reference parameter to eliminate the influence of environmental characteristic deviations under different office environments on the Mel-Cepstral Coefficient Sequence. When calculating the Euclidean distance, since each frequency domain component of the input speech frame spectrum distribution vector and the environmental reference spectrum distribution vector has been pre-converted into a logarithmic power spectrum scalar in decibels, the multidimensional Euclidean distance between the two in high-dimensional geometric space reflects the comprehensive logarithmic value of the power spectrum energy deviation across the entire frequency band. By performing one-dimensional mean calibration on the spatial distance, its scalar calculation result can be directly compared with the decibel logarithmic threshold in terms of dimensional equivalence. When the calculated geometric distance value exceeds 3.5 dB, it means that the current environmental noise floor has undergone acoustic shift.

[0042] Feature tensor builder via compensation operator Correcting the timing discontinuity caused by sampling step size switching, assuming the sampling step size is changed from... Switch to Furthermore, the switching point generates a timing offset. Feature tensor builder based on Displacement compensation is performed in the characteristic tensor, and the calculation logic is as follows: ,in, To compensate for the number of frames, For the current sampling period, As a rounding function, when constructing the temporal acoustic feature tensor, the system will... Each feature frame is temporally shifted, and a weighted average is applied to the overlapping regions after the shift. This process eliminates feature breaks and ensures that the tensors received by the decoding model have continuous physical properties on the time axis. The system sets the upper limit of the sampling step size switching frequency to 5Hz, which is limited by setting a minimum forced locking time of 200ms after the switch to prevent the feature stream from oscillating repeatedly in a short time. The closed-loop logic constructed in this way ensures the stability of the feature output under different signal-to-noise ratios.

[0043] The above description is only a few preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with the technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A speech recognition method for voice input in expense reimbursement, characterized in that, Includes the following steps: Step S101: The acquired continuous speech signal is buffered into a data buffer to generate a digital audio stream; Step S102: Perform short-time Fourier transform on the digitized audio stream, extract the multidimensional Mel-Cepstral coefficient feature sequence, construct the reference acoustic feature frame stream and store it in the storage array; Step S103: Read adjacent reference acoustic feature frames and calculate the acoustic energy difference and frequency domain spectral flux variability between adjacent reference acoustic feature frames. When the acoustic energy difference exceeds the preset energy threshold, or the frequency domain spectral flux variability exceeds the preset variability threshold, the corresponding reference acoustic feature frame is marked as a candidate entity phoneme anchoring frame on the time axis, and its timestamp is written into the control register. Step S104: Read the timestamp difference between adjacent candidate entity phoneme anchoring frames in the control register, and adjust the feature sampling step size according to the geometric distribution density reflected by the timestamp difference: In the high-density target interval where the temporal interval between two adjacent candidate entity phoneme anchoring frames is less than the preset time window, reduce the feature sampling step size to the first step size; In the low-density redundant interval where there are no unmarked candidate entity phoneme anchoring frames, expand the feature sampling step size to the second step size, where the first step size is smaller than the second step size. Step S105 involves concatenating the acoustic feature frames extracted by different feature sampling steps in their original temporal sequence to construct a discrete temporal acoustic feature tensor, which is then input into the acoustic decoding model to complete the alignment decoding and output the reimbursement text data.

2. The speech recognition method for expense reimbursement voice input according to claim 1, characterized in that, While extracting the multidimensional Mel-Cepstral coefficient feature sequence by performing short-time Fourier transform on the digitized audio stream, the transient signal-to-noise ratio (SNR) of the digitized audio stream across the entire frequency domain is monitored in real time. When the transient SNR drops by more than 15 dB within 10 ms, it is determined that there is a sudden non-speech parasitic noise in the digitized audio stream. The marking action of anchoring candidate entity phonemes is then suspended, and the feature sampling step size is controlled to remain unchanged at the current value to prevent oversampling caused by noise pseudo-energy, until the transient SNR of the entire frequency domain recovers to above the safe threshold.

3. The speech recognition method for expense reimbursement voice input according to claim 1, characterized in that, In step S103, when calculating the acoustic energy difference value, parameter D is configured as the absolute value of the difference between the short-time logarithmic energy value of the reference acoustic feature frame at the current position and the short-time logarithmic energy value of the reference acoustic feature frame at the previous adjacent position; when the acoustic energy difference value exceeds the preset energy threshold, the reference acoustic feature frame at the corresponding position is marked as a candidate entity phoneme anchoring frame.

4. A speech recognition method for voice input of expense reimbursement according to claim 1, characterized in that, Step S104 further includes the following sub-steps: Step S1041, using the counter register to accumulate the duration of continuous use of the second step; Step S1042, when the duration exceeds 300ms, controlling the feature sampling step size to be downsampled from the second step size back to a fixed value of 15ms until the next candidate entity phoneme anchor frame that meets the threshold condition is detected.

5. A voice recognition method for voice input of expense reimbursement according to claim 1, characterized in that, The preset time window is 50ms, the first step is 5ms long, and the second step is 30ms long. In the low-density interval of the unlabeled candidate entity phoneme anchor frame, sparse sampling is implemented through the second step to discard the background noise of the environment.

6. A speech recognition method for voice input of expense reimbursement according to claim 1, characterized in that, Step S105 further includes the following sub-steps: Step S1051, obtaining the first acoustic feature frame sequence extracted using the first step length within the high-density target interval, and obtaining the second acoustic feature frame sequence extracted using the second step length within the low-density redundant interval; Step S1052, extracting the original timestamps corresponding to each first acoustic feature frame sequence and each second acoustic feature frame sequence; Step S1053, concatenating and splicing the first acoustic feature frame sequence and the second acoustic feature frame sequence according to the original timestamps from early to late to construct a discrete temporal acoustic feature tensor, so as to reduce the size of the feature matrix input to the acoustic decoding model.

7. A voice recognition method for voice input of expense reimbursement according to claim 1, characterized in that, Step S103, calculating the frequency domain spectral flux variability, further includes the following sub-steps: Step S1031, obtaining the frequency domain spectral distribution vector of the reference acoustic feature frame at the current position on the time axis, and obtaining the frequency domain spectral distribution vector of the reference acoustic feature frame at the previous adjacent position; Step S1032, calculating the energy change of the frequency domain spectral distribution vector at the current position and the frequency domain spectral distribution vector at the previous adjacent position in each frequency domain channel; Step S1033, summing the squares of the energy changes in each frequency domain channel to generate the frequency domain spectral flux variability.

8. A speech recognition method for voice input of expense reimbursement according to claim 1, characterized in that, The storage array uses a dual-port static random access memory array; a reference acoustic feature frame stream is cyclically written to the storage array through the first port, and adjacent reference acoustic feature frames are periodically read from the storage array through the second port.

9. A speech recognition method for voice input of expense reimbursement according to claim 1, characterized in that, In step S104, the feature sampling step size is adjusted based on the timestamp difference between adjacent candidate entity phoneme anchor frames in the control register, including: when the timestamp difference between two adjacent candidate entity phoneme anchor frames is less than a preset time window, the current feature sampling step size is switched to the first step size.

10. A speech recognition method for voice input of expense reimbursement according to claim 1, characterized in that, In step S105, the acoustic decoding model includes a deep neural network architecture and a connectionist temporal classification decoder; the discrete temporal acoustic feature tensor is derived using the acoustic decoding model, and the corresponding reimbursement text data is output.

Images