Methods and Systems for Extracting Key Information from Fire Alarm Voice Recordings

Abstract

This invention discloses a method and system for extracting key information from fire alarm voice messages, comprising: performing frame-by-frame parsing processing on the raw voice time-series waveform data; determining the type of noise interference and marking regions based on time-series delay feature data; marking waveform regions without fixed time delay and without repeating homogeneous waveforms as alarm regions; generating voice integrity time-series verification data; dynamically generating dual-weighted dynamic adjustment data based on the time-series delay feature data and the voice integrity time-series verification data; performing graded noise reduction processing step by step according to the dual-weighted dynamic adjustment data; and finally outputting valid alarm voice messages for recognition. This method minimizes the possibility that the alarm voice of the person receiving the alarm may be obscured and cannot be effectively extracted, greatly improving the accuracy of key information extraction. This invention features high accuracy in key information extraction and strong alarm voice recognition capabilities.

Description

Technical Field

[0001] This invention relates to the field of emergency call voice recognition technology, specifically to a method and system for extracting key information from fire alarm voice messages. Background Technology

[0002] Real fire environments are complex and ever-changing. Key alarm elements such as the location of the fire, the status of trapped personnel, the type of fire, the operating status of on-site equipment, and information on special hazards rely entirely on the on-site verbal communication of the person reporting the fire. As a core pre-processing technology that assists fire communicators in intelligently extracting fire information and rapidly dispatching rescue efforts, the recognition accuracy of the voice key information extraction system directly determines the accuracy of fire location and the timeliness of rescue. Once information is missed, misdetected, or confused, it can easily lead to serious problems such as deviations in rescue points, improper allocation of rescue forces, and lack of high-risk prediction, directly threatening the safety of personnel and the safety of fire rescue operations.

[0003] After receiving an emergency call, fire communication operators will inquire about the fire situation. When the caller is in an enclosed space such as a corridor, various types of noise exist, creating a complex environment of speech interference that easily obscures crucial information in the voice. Even though existing fire alarm voice noise reduction and recognition technologies can handle environments with multiple speech interferences, the types of noise interference in a fire scenario far exceed the processing range of the information extraction system. In particular, callers are prone to coughing in dense smoke environments. The voice generated by the caller's normal speech and the coughing impulse noise each independently form corresponding echo components. The original sound source waveform and the corresponding echo waveform continuously overlap and intertwine in the time domain, severely interfering with the effective acquisition of crucial voice information by the fire communication operator, greatly reducing the efficiency of alarm dispatch. Therefore, it is necessary to design a method and system for extracting crucial information from fire alarm voice calls with high accuracy and strong alarm voice recognition capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for extracting key information from fire alarm voice messages, in order to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method and system for extracting key information from fire alarm reception voice messages, comprising:

[0006] In response to a fire alarm call request, the system continuously collects the original voice within the call request, generates raw voice time-series waveform data, performs frame-by-frame parsing on the raw voice time-series waveform data, compares the waveform similarity and waveform occurrence time difference frame by frame, and extracts time-series delay feature data. The time-series delay feature data includes the time difference value of adjacent similar waveforms, the frequency of waveform delay occurrence, and the distribution density of repeated waveforms with fixed time delay.

[0007] Based on the timing delay feature data, noise interference type determination and area marking are completed. Waveform areas where the detected waveform has no fixed time delay and no repeating same source waveforms are marked as alarm areas, and speech integrity timing verification data is generated based on the current waveform area segment.

[0008] Based on the timing delay feature data and the speech integrity timing verification data, dual-weight dynamic adjustment data is dynamically generated, and hierarchical noise reduction processing is performed step by step according to the dual-weight dynamic adjustment data.

[0009] The noise-reduced voice signal is collected, a mode switching command is executed, noise-reduced output real-time feedback time-series data is generated, dual-weight dynamic adjustment data is iteratively fine-tuned, and hierarchical noise reduction processing is performed repeatedly until the target alarm voice that meets the preset recognition information is extracted. Based on the target alarm voice, key fire information is extracted.

[0010] According to the above technical solution, the frame-by-frame comparison of waveform similarity and waveform occurrence time difference extracts timing delay feature data. This timing delay feature data includes the time difference between adjacent similar waveforms, the frequency of waveform delay occurrence, and the distribution density of repeated waveforms with fixed time delays, including:

[0011] The machine divides the raw audio time-series waveform sampling sequence into frames according to a preset duration. Each frame generates an independent frame waveform sampling sequence, and all frame waveform sampling sequences form an ordered frame queue, recording the absolute timestamp corresponding to each frame.

[0012] Traverse the frame queue, take the current frame waveform sampling sequence in turn, and perform point-to-point amplitude squared difference operation with all historical frame waveform sampling sequences in the queue to generate inter-frame similarity values.

[0013] Filter N groups of frames whose inter-frame squared difference values ​​are lower than the preset squared difference value threshold, retrieve the absolute timestamps bound to each of the N groups of frames, and subtract the timestamp of the previous frame from the timestamp of the later frame to obtain the time difference values ​​of the N groups of the same waveform, where N is the rated number of frames retrieved within the preset unit time.

[0014] The three types of parameters included in the generation of time-series delay feature data are:

[0015] The time difference between adjacent similar waveforms is the set of time difference values ​​corresponding to all successfully matched adjacent highly similar waveforms.

[0016] Waveform delay frequency is the total number of times the same time difference is repeated.

[0017] The distribution density of fixed-delay repetitive waveforms is the percentage of frames with fixed delay and repetitive waveforms within the preset unit of time.

[0018] The three types of parameters are combined into a structured dataset.

[0019] According to the above technical solution, the step of marking a waveform region with no fixed time delay and no repeating homologous waveforms as an alarm region, and generating voice integrity timing verification data based on the current waveform region segment, includes:

[0020] Traverse all frames of the original audio, filter out the audio frames corresponding to the alarm area, and reassemble them according to the order of collection time to form an ordered sequence of human voice frames.

[0021] The method for determining whether there are gaps in the human voice punctuation within a single frame of the ordered human voice frame sequence includes:

[0022] Extract the amplitude values ​​of all sampling points within a single frame and compare them with a preset threshold for determining the amplitude of human voice interruption. Determine the sampling interval where the amplitude values ​​of multiple consecutive sampling points within the frame are lower than the interruption threshold as a sound gap within the single frame. Record the frame number corresponding to the gap. Mark the start and end sampling times within the frame as a time-series interruption point group. Arrange all the identified time-series interruption point groups in chronological order to obtain the time-series interruption point distribution data.

[0023] The effective frame ratio of human voice is calculated statistically, and the total number of complete frames of the original voice and the total number of frames marked as effective alarm personnel voice regions are obtained. The effective frame ratio of human voice is obtained by dividing the total number of frames of the effective alarm personnel voice regions by the total number of complete frames.

[0024] Calculate the length of the coughing and choking interruption interval corresponding to each of the time-series discontinuity point groups, retrieve the sampling start time stamp and sampling end time stamp, and subtract the sampling start time stamp from the sampling end time stamp to obtain the duration of a single interruption; traverse all the time-series discontinuity points, record the duration of each single interruption, and summarize the entire duration set as the coughing and choking interruption interval length data.

[0025] The temporal discontinuity point distribution data, the proportion of effective human voice frames, and the length of coughing and choking intervals are respectively output as three rows of sub-data for speech integrity temporal verification data. The three rows of sub-data share the frame timestamp and frame number of the original speech as a unified index.

[0026] According to the above technical solution, the step of performing hierarchical noise reduction processing based on the dual-weighted dynamic adjustment data includes:

[0027] Based on the timing verification data of the voice integrity, the timing discontinuity point groups within the alarm area are located. Combined with the timing delay feature data, residual echo pseudo-waveforms and pulse interference noise between adjacent timing discontinuity point groups are removed. The human voice timing waveform features and frequency band distribution features between adjacent timing discontinuity point groups are extracted as repair benchmark feature data. Based on the repair benchmark feature data, the timing waveform smoothing and completion processing is performed on the discontinuous voice information chain, and the timing connection calibration processing is performed on the discontinuously split voice information chain.

[0028] According to the above technical solution, the step of determining the noise interference type and marking the region based on the time delay feature data includes:

[0029] If a waveform with a fixed time delay is detected and repeated waveforms of the same source appear a preset threshold number of times, it is determined to be multiple echo interference in the corridor, and the echo time delay value and repetition density are counted.

[0030] According to the above technical solution, the dual-weight dynamic adjustment data also has graded noise reduction weights corresponding to multiple echoes. After detecting that the frequency of waveform delay and the distribution density of fixed-delay repetitive waveforms are both higher than the threshold based on the time delay feature data, the graded noise reduction weights corresponding to multiple echoes are increased to fix the highest noise reduction priority of multiple echoes. At the same time, the global voice integrity constraint weights are simultaneously increased. The remaining noises are matched with corresponding graded noise reduction weights according to their respective detected time fluctuation intensity. Graded noise reduction processing is performed step by step according to the dual-weight dynamic adjustment data. Based on the increased graded noise reduction weights of multiple echoes, the delayed and repetitive reflection waveforms are targeted and stripped, while retaining the voice time discontinuity and coughing auxiliary waveform features in the alarm area.

[0031] According to the above technical solution, the process of acquiring the noise-reduced speech signal, executing a mode switching command, and generating real-time feedback time-series data for noise-reduced output includes:

[0032] By continuously comparing the overall distribution offset of waveform time-series distortion feature parameters of adjacent frames, synchronous abrupt change nodes of multi-dimensional waveform distortion features are captured. Based on the consistency of abrupt changes in the overall waveform distortion distribution, the dynamic switching behavior of the fire scene is determined and a mode switching command is generated.

[0033] According to the above technical solution, the step of extracting key fire information based on the target alarm voice includes:

[0034] Based on the frequency of waveform delay occurrence, the number of echoes corresponding to different propagation directions is distinguished. Based on the number of echoes in different directions, the distance between the alarm person's sound source and the corresponding wall is obtained, and the spatial dimensions of each direction in the corridor are determined. At the same time, based on the time difference of the adjacent similar waveforms in the same vertical direction and the distribution density of the fixed-delay repeated waveforms, the farthest distance from the alarm person's current position to the floor below is obtained.

[0035] According to the above technical solution, the fire alarm voice key information extraction system includes:

[0036] The data acquisition module is used to respond to fire alarm call requests, continuously acquire the original voice within the call request, generate original voice time-series waveform data, perform frame-by-frame parsing processing on the original voice time-series waveform data, compare waveform similarity and waveform occurrence time difference frame by frame, and extract time-series delay feature data. The time-series delay feature data includes the time difference value of adjacent similar waveforms, waveform delay occurrence frequency, and distribution density of fixed-delay repeating waveforms.

[0037] The noise determination module is used to determine the type of noise interference and mark the region based on the timing delay feature data. It marks the waveform region where the detected waveform has no fixed time delay and no repeating homologous waveforms as an alarm region, and generates voice integrity timing verification data based on the current waveform region segment.

[0038] The weight adjustment module is used to dynamically generate dual-weight dynamic adjustment data based on the timing delay feature data and the speech integrity timing verification data, and to perform hierarchical noise reduction processing step by step according to the dual-weight dynamic adjustment data.

[0039] The key information extraction module is used to execute mode switching instructions, generate noise reduction output real-time feedback time-series data, compare the noise reduction output real-time feedback time-series data with preset standard alarm voice time-series features, iteratively fine-tune the dual-weight dynamic adjustment data based on the comparison results, and cyclically perform graded noise reduction processing until a target alarm voice that meets the preset recognition information is extracted, and extract key fire information based on the target alarm voice.

[0040] According to the above technical solution, a storage medium stores a computer program thereon, and when the computer program is executed by a processor, it implements the method for extracting key information from fire alarm voice as described in the above technical solution.

[0041] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention performs frame-by-frame parsing on the raw data of the speech time-series waveform, and completes noise interference type determination and region marking based on time-series delay feature data. When a waveform region with no fixed time delay and no repeating homologous waveforms is detected, it is marked as an alarm region. Speech integrity time-series verification data is generated. Based on the time-series delay feature data and the speech integrity time-series verification data, dual-weight dynamic adjustment data is dynamically generated. Based on the dual-weight dynamic adjustment data, hierarchical noise reduction processing is performed step by step, and finally, effective alarm voice is output for recognition. This minimizes the possibility that the alarm voice of the alarm personnel is obscured and cannot be effectively extracted, and greatly improves the accuracy of key information extraction. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a flowchart of a method for extracting key information from fire alarm voice recordings provided in an embodiment of the present invention.

[0044] Figure 2 This is a block diagram of a fire alarm voice key information extraction system provided in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of a fire scene illustrating the method for extracting key information from fire alarm voice recordings provided in an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1 , 3 The flowchart of the method and system for extracting key information from fire alarm voice recordings provided in the embodiments of the present invention, and the fire scenario diagram of the method for extracting key information from fire alarm voice recordings, are shown below. Figure 1 , 3 It can be seen that the method and system for extracting key information from fire alarm voice messages include:

[0048] Step S1: Respond to the fire alarm call request, continuously collect the original voice in the call request, generate the original voice time-series waveform data, perform frame-by-frame parsing processing on the original voice time-series waveform data, compare the waveform similarity and waveform occurrence time difference frame by frame, and extract the time-series delay feature data. The time-series delay feature data includes the time difference value of adjacent similar waveforms, the frequency of waveform delay occurrence, and the distribution density of fixed-delay repeating waveforms.

[0049] Step S2: Based on the time-series delay feature data, determine the type of noise interference and mark the region. Mark waveform regions with fixed time delays and repeating homologous waveforms as single-echo coverage areas. Single echoes have little impact on the key voice information of the alarm personnel and will not affect the noise in the alarm area; therefore, adaptive echo cancellation, time-domain template matching subtraction filtering, and other techniques can be used. Mark waveform regions without fixed time delays and without repeating homologous waveforms as alarm areas. If a waveform with a fixed time delay is detected and occurs a preset threshold number of times... Repeated echoes from the same source are identified as multiple echoes in the corridor, and the echo delay value and repetition density are calculated. If a waveform with a fixed delay is detected and repeated echoes from the same source occur a preset threshold number of times, it is identified as multiple echoes in the corridor, and the echo delay value and repetition density are calculated. Based on the current waveform region segment, speech integrity timing verification data is generated, which is mainly used for preprocessing to filter out several types of echoes. A single echo is determined only by the time difference of adjacent similar waveforms to determine whether the similar waveform is a reflection echo from the corridor. If the echo is a single echo, there will be no situation where there are echoes from the same source but no fixed delay.

[0050] Step S3: Based on the time-series delay feature data and the speech integrity time-series verification data, dynamic dual-weight dynamic adjustment data is generated. The dual-weight dynamic adjustment data is set with global voice integrity constraint weight, graded noise reduction weight, and graded noise reduction weight corresponding to multiple echoes. Graded noise reduction processing is performed step by step according to the dual-weight dynamic adjustment data. It is mainly used to adjust the weights to eliminate noise. After the frequency of waveform delay and the distribution density of fixed-delay repetitive waveforms are both higher than the threshold based on the time-series delay feature data, the graded noise reduction weight corresponding to multiple echoes is increased to fix the highest noise reduction priority of multiple echoes. At the same time, the global voice integrity constraint weight is increased simultaneously. The remaining noise is matched with the corresponding graded noise reduction weight according to the intensity of their respective detected time-series fluctuations. Graded noise reduction processing is performed step by step according to the dual-weight dynamic adjustment data. Based on the increased graded noise reduction weight of multiple echoes, the delayed and repetitive reflection waveforms are targeted and stripped, while retaining the speech time-series interruption and coughing auxiliary waveform features in the alarm area.

[0051] Set the single echo detection threshold: waveform delay frequency < 3 times; multiple echo detection threshold: waveform delay frequency ≥ 3 times, downlink distribution density ≥ 0.75; cough echo detection threshold: attached cough waveform, delay range 70-95ms, same source delay waveform density ≥ 0.70; native cough protection threshold: no delay, no delay replica waveform, judged as native effective cough waveform, noise reduction and suppression are prohibited.

[0052] Four categories of benchmark-based noise reduction weights are used as the baseline value for silent, echo-free environments. The smaller the protection benchmark weight value, the weaker the noise reduction suppression and the stronger the protection: Original cough and choking protection benchmark weight W c0 =0.20; Single-echo baseline noise reduction weight W s0 =0.40; Multi-echo reference noise reduction weight W m0 =0.45; Coughing-derived echo baseline noise reduction weight W k0 =0.5.

[0053] Combined with feature overshoot factor linkage weighting, the weight values ​​range from 0 to 1. The larger the value, the stronger the noise reduction and stripping effect. The single-echo dynamic weighting formula is as follows: Where X is the echo frequency in one direction; the dynamic weighting formula for multiple echoes is: Where α is the preset echo up-adjustment coefficient, and ρ is the preset downlink waveform distribution density; the dynamic weighting formula for cough echo is: Where β is the preset cough echo upregulation coefficient; native cough fixed protection weight: The waveform remains fixed throughout the entire process, without dynamic adjustment, and the original coughing waveform is permanently preserved.

[0054] Single-echo weighting calculation (lateral frequency X=4): Moderate noise reduction to weaken single wall reflection echoes; multiple echo weighting calculation (ρ=0.91, α=1.6): Highest noise reduction priority across the entire domain, effectively stripping away echoes from multiple floors; coughing echo weighting calculation (compliance trigger coefficient β=1.4): Specialized stripping of delayed floor-reflected cough echoes generated by the original coughing sensation, without damaging the original coughing waveform; fixed protection weight for the original alarm coughing sensation: It features fixed protection throughout the entire process and does not perform noise reduction processing on the original coughing waveform without time delay.

[0055] The noise reduction priority is determined according to the weight value. The higher the weight, the higher the noise reduction priority and the stronger the noise reduction stripping effect. The priority is sorted from high to low as follows: cough echo weight 0.7 > multiple echo weight 0.655 > single echo weight 0.56 > original cough protection weight 0.2.

[0056] If the subsequent real-time waveform delay frequency N < 3 and the downlink distribution density ρ < 0.75, it is determined that the multiple echoes have dissipated. m Returning to the baseline value of 0.45; if the cough echo delay deviates from the 70-95ms range, W k It returns to the baseline value of 0.50; the single echo weight changes in real time with the real-time echo frequency N. The higher the frequency, the slightly better the single echo noise reduction; the native cough weight is set to be fixed and is not affected by echo characteristics.

[0057] Step S4: Acquire the noise-reduced speech signal, execute the mode switching command, generate noise-reduced output real-time feedback timing data, compare the noise-reduced output real-time feedback timing data with the preset standard alarm speech timing features, iteratively fine-tune the dual-weight dynamic adjustment data based on the comparison results, and repeatedly perform graded noise reduction processing until the target alarm speech that meets the preset recognition information is extracted. Extract key fire information based on the target alarm speech, greatly improving the accuracy of key information extraction, and is mainly used for key information extraction.

[0058] In embodiments of the present invention, frame-by-frame parsing is performed on the raw voice time-series waveform data. Based on time-series delay feature data, noise interference type determination and region marking are completed. When a waveform region with no fixed time delay and no repeating homologous waveforms is detected, it is marked as an alarm region. Voice integrity time-series verification data is generated. Based on the time-series delay feature data and the voice integrity time-series verification data, dual-weight dynamic adjustment data is dynamically generated. Based on the dual-weight dynamic adjustment data, hierarchical noise reduction processing is performed step by step, and finally, effective alarm voice is output for recognition. This minimizes the possibility that the alarm voice of the alarm personnel is obscured and cannot be effectively extracted, and greatly improves the accuracy of key information extraction.

[0059] In this embodiment, the frame-by-frame comparison of waveform similarity and waveform occurrence time difference extracts timing delay feature data. This timing delay feature data includes the time difference between adjacent similar waveforms, the frequency of waveform delay occurrence, and the distribution density of repeating waveforms with fixed time delays, including:

[0060] The machine divides the raw audio time-series waveform sampling sequence into frames according to a preset duration. Each frame generates an independent frame waveform sampling sequence, and all frame waveform sampling sequences form an ordered frame queue, recording the absolute timestamp corresponding to each frame.

[0061] Traverse the frame queue, take the waveform sampling sequence of the current frame in turn, and perform point-to-point amplitude squared difference operation with the waveform sampling sequences of all historical frames in the queue. The smaller the squared difference, the higher the similarity between the two frames. Generate an inter-frame similarity value.

[0062] Calculate the time difference between waveforms in highly similar frames, filter N groups of frames whose inter-frame squared difference values ​​are lower than a preset squared difference value threshold, retrieve the absolute timestamps bound to each of the N groups of frames, subtract the timestamp of the previous frame from the timestamp of the later frame, and obtain the time difference values ​​of the N groups of homogeneous waveforms. The homogeneous waveforms will match the same or the time difference value within the preset threshold, where N is the rated number of frames retrieved within a preset unit time.

[0063] The three types of parameters included in the generation of time-series delay feature data are:

[0064] The time difference between adjacent similar waveforms is the set of time difference values ​​corresponding to all successfully matched adjacent highly similar waveforms.

[0065] Waveform delay frequency is the total number of times the same time difference is repeated.

[0066] The distribution density of fixed-delay repetitive waveforms is the percentage of frames with fixed delay and repetitive waveforms within the preset unit of time.

[0067] The three types of parameters are combined into a structured dataset.

[0068] A structured noise feature dataset is generated. Various types of conventional noise feature data are obtained by separating the effective human voice waveform and multiple echo waveforms from the original speech time-series waveform, resulting in pure interference waveform data. The data dimensions cover the time-series distribution range, instantaneous energy amplitude, and inter-frame energy fluctuation parameters of various noise types. The structured noise feature dataset completes independent attribute calibration based on the time-series delay feature data, establishing a visual correspondence between noise feature parameters and graded noise reduction weights. This dataset initially assists in accurately distinguishing noise and human voice distortion ranges, avoiding feature aliasing recognition errors, and later integrates with dual-weight dynamic adjustment logic.

[0069] In this embodiment, the step of marking a waveform region with no fixed time delay and no repeating homologous waveforms as an alarm region, and generating voice integrity timing verification data based on the current waveform region segment, includes:

[0070] Traverse all frames of the original audio, filter out the audio frames corresponding to the alarm area, and reassemble them according to the order of collection time to form an ordered sequence of human voice frames.

[0071] The method for determining whether there are gaps in the human voice punctuation within a single frame of the ordered human voice frame sequence includes:

[0072] Extract the amplitude values ​​of all sampling points within a single frame and compare them with a preset threshold for determining the amplitude of human voice interruption. Determine the sampling interval where the amplitude values ​​of multiple consecutive sampling points within the frame are lower than the interruption threshold as a sound gap within the single frame. Record the frame number corresponding to the gap. Mark the start and end sampling times within the frame as a time-series interruption point group. Arrange all the identified time-series interruption point groups in chronological order to obtain the time-series interruption point distribution data.

[0073] The effective frame ratio of human voice is calculated statistically, and the total number of complete frames of the original voice and the total number of frames marked as effective alarm personnel voice regions are obtained. The effective frame ratio of human voice is obtained by dividing the total number of frames of the effective alarm personnel voice regions by the total number of complete frames.

[0074] Calculate the length of the coughing and choking interruption interval corresponding to each of the time-series discontinuity point groups, retrieve the sampling start time stamp and sampling end time stamp, and subtract the sampling start time stamp from the sampling end time stamp to obtain the duration of a single interruption; traverse all the time-series discontinuity points, record the duration of each single interruption, and summarize the entire duration set as the coughing and choking interruption interval length data.

[0075] The temporal discontinuity point distribution data, the proportion of effective human voice frames, and the length of coughing and choking intervals are respectively output as three rows of sub-data for speech integrity temporal verification data. The three rows of sub-data share the frame timestamp and frame number of the original speech as a unified index.

[0076] In this embodiment, the step of performing hierarchical noise reduction processing based on the dual-weighted dynamic adjustment data includes:

[0077] Based on the timing verification data of the voice integrity, the timing discontinuity point groups within the alarm area are located. Combined with the timing delay feature data, residual echo pseudo-waveforms and pulse interference noise between adjacent timing discontinuity point groups are removed. The human voice timing waveform features and frequency band distribution features between adjacent timing discontinuity point groups are extracted as repair benchmark feature data. Based on the repair benchmark feature data, the timing waveform smoothing and completion processing is performed on the discontinuous voice information chain, and the timing connection calibration processing is performed on the discontinuously split voice information chain.

[0078] In this embodiment, the process of acquiring the noise-reduced speech signal, executing a mode switching command, and generating real-time feedback timing data for noise-reduced output includes:

[0079] Multiple fire scene switching recognition, acquire the original speech time-domain waveform data of continuous time sequence during the fire alarm call, perform time-frequency conversion processing on the original speech time-domain waveform data, and extract waveform time-series distortion feature parameters frame by frame. The waveform time-series distortion feature parameters include at least waveform time-domain overlap, waveform high-frequency attenuation, and waveform pulse distortion, which correspond to the inherent waveform features of corridor reverberation distortion, sprinkler water mist high-frequency loss, and dense smoke cough pulse distortion, respectively.

[0080] By continuously comparing the overall distribution offset of waveform time-series distortion feature parameters of adjacent frames, synchronous abrupt change nodes of multi-dimensional waveform distortion features are captured. Based on the consistency of abrupt changes in the overall waveform distortion distribution, the dynamic switching behavior of the fire scene is determined and a mode switching command is generated.

[0081] Based on the dominant distortion type of waveform timing distortion characteristic parameters before and after switching, the corresponding dedicated speech correction processing logic is adaptively matched. The dedicated speech correction processing logic includes reverberation overlap waveform stripping logic, high frequency loss waveform compensation logic, pulse distortion waveform smoothing logic, and multi-distortion coupled waveform layered decoupling logic.

[0082] By matching the dedicated voice correction processing logic, the real-time raw voice time-domain waveform data is targeted to correct distortion, eliminating the interference of waveform distortion abrupt changes caused by scene switching, and outputting standardized and clean voice data with stable timing.

[0083] Firefighting semantic feature parsing is performed on the standardized clean voice data to achieve stable identification and output of key fire alarm information under dynamic scene switching.

[0084] In this embodiment, extracting key fire information based on the target alarm voice includes:

[0085] Based on the frequency of waveform delay occurrence, the number of echoes corresponding to different propagation directions is distinguished. Based on the number of echoes in different directions, the distance between the alarm person's sound source and the corresponding wall is obtained, and the spatial dimensions of each direction in the corridor are determined. At the same time, based on the time difference of the adjacent similar waveforms in the same vertical direction and the distribution density of the fixed-delay repeated waveforms, the farthest distance from the alarm person's current position to the floor below is obtained.

[0086] Due to the influence of the escape posture in the stairwell, and the fact that when the person making the alarm is generally inside the stairwell, their voice is directed downwards as they need to run out of the floor. Voices and coughing noises primarily propagate vertically downwards, and secondarily horizontally towards the surrounding walls. Upward sound waves attenuate rapidly, resulting in no effective upward echo signal. Based on echo delay and the number of floors below, a four-level classification is applied. If the detected delay is ≤30ms, it is assumed that the floor below the alarm person is ≤1 floor, the stairwell is narrow, and the search and rescue difficulty is low, thus classifying it as Level 1 (High Rescue Difficulty). If the detected delay... If the delay is 30ms-60ms, the fire is presumably on the 2nd-3rd floor below the person who triggered the alarm. The fire is under control and can be carried out by conventional downward search and rescue, and the level is classified as Level 2, high rescue difficulty. If the delay is 60ms-100ms, the fire is presumably on the 4th-5th floor below the person who triggered the alarm. The stairwell is open and there is significant backflow of smoke, making downward search and rescue risky, and the level is classified as Level 3, high rescue difficulty. If the delay is ≥100ms, the fire is presumably on the 6th floor below the person who triggered the alarm. The fire is widespread and downward search and rescue operations are prohibited, and the level is classified as Level 4, high rescue difficulty.

[0087] The waveform delay frequency detected was 4 times in the horizontal left and right wall directions and 7 times in the horizontal depth wall direction. The frequency of the waveform delay in the horizontal depth wall direction was greater than that in the horizontal left and right wall directions, indicating that the corridor type was a long and narrow corridor. The adjacent similar waveform delay in the vertical downward floor direction was 88ms, and the distribution density of the fixed delay repeating waveform was 0.91. The length below the alarm person cannot be accurately determined by the delay magnitude alone. Noise in non-vertical directions is easily blocked by walls, debris, windows, etc., and the sound wave stops reflecting when it hits an obstacle. The predicted physical distance deviates from the theoretical distance of the delay, leading to misjudgment. Therefore, the distribution density of the delay repeating waveform is close to the full value, which can verify that this space is unobstructed, without partitions or blockages, and the sound wave can reflect back and forth repeatedly. Based on the waveform information contained in the obtained time delay feature data, the spread of the fire on the floor where the alarm person is located can be predicted.

[0088] Please see Figure 2 Based on the same concept as the above embodiments, embodiments of the present invention also provide a fire alarm voice key information extraction system, including:

[0089] The data acquisition module is used to respond to fire alarm call requests, continuously acquire the original voice within the call request, generate original voice time-series waveform data, perform frame-by-frame parsing processing on the original voice time-series waveform data, compare waveform similarity and waveform occurrence time difference frame by frame, and extract time-series delay feature data. The time-series delay feature data includes the time difference value of adjacent similar waveforms, waveform delay occurrence frequency, and distribution density of fixed-delay repeating waveforms.

[0090] The noise determination module is used to determine the type of noise interference and mark the region based on the timing delay feature data. The waveform region where the detected waveform has a fixed time delay and has repeating same source waveforms is marked as an echo coverage region. The waveform region where the detected waveform has no fixed time delay and no repeating same source waveforms is marked as an alarm region. The module also generates voice integrity timing verification data based on the current waveform region segment.

[0091] The weight adjustment module is used to dynamically generate dual-weight dynamic adjustment data based on the timing delay feature data and the speech integrity timing verification data, and to perform hierarchical noise reduction processing step by step according to the dual-weight dynamic adjustment data.

[0092] The key information extraction module is used to execute mode switching instructions, generate noise reduction output real-time feedback time-series data, compare the noise reduction output real-time feedback time-series data with preset standard alarm voice time-series features, iteratively fine-tune the dual-weight dynamic adjustment data based on the comparison results, and cyclically perform graded noise reduction processing until a target alarm voice that meets the preset recognition information is extracted, and extract key fire information based on the target alarm voice.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting key information from fire alarm voice messages, characterized in that: include: In response to a fire alarm call request, the system continuously collects the original voice within the call request, generates raw voice time-series waveform data, performs frame-by-frame parsing on the raw voice time-series waveform data, compares the waveform similarity and waveform occurrence time difference frame by frame, and extracts time-series delay feature data. The time-series delay feature data includes the time difference value of adjacent similar waveforms, the frequency of waveform delay occurrence, and the distribution density of repeated waveforms with fixed time delay. Based on the timing delay feature data, noise interference type determination and area marking are completed. Waveform areas where the detected waveform has no fixed time delay and no repeating same source waveforms are marked as alarm areas, and speech integrity timing verification data is generated based on the current waveform area segment. Based on the timing delay feature data and the speech integrity timing verification data, dual-weight dynamic adjustment data is dynamically generated, and hierarchical noise reduction processing is performed step by step according to the dual-weight dynamic adjustment data. The noise-reduced voice signal is collected, a mode switching command is executed, noise-reduced output real-time feedback time-series data is generated, dual-weight dynamic adjustment data is iteratively fine-tuned, and hierarchical noise reduction processing is performed repeatedly until the target alarm voice that meets the preset recognition information is extracted. Based on the target alarm voice, key fire information is extracted.

2. The method for extracting key information from fire alarm voice messages according to claim 1, characterized in that: The frame-by-frame comparison of waveform similarity and waveform occurrence time difference extracts temporal delay feature data. This temporal delay feature data includes the time difference between adjacent similar waveforms, the frequency of waveform delay occurrence, and the distribution density of repeated waveforms with fixed time delays. The machine divides the raw audio time-series waveform sampling sequence into frames according to a preset duration. Each frame generates an independent frame waveform sampling sequence, and all frame waveform sampling sequences form an ordered frame queue, recording the absolute timestamp corresponding to each frame. Traverse the frame queue, take the current frame waveform sampling sequence in turn, and perform point-to-point amplitude squared difference operation with all historical frame waveform sampling sequences in the queue to generate inter-frame similarity values. Filter N groups of frames whose inter-frame squared difference values ​​are lower than the preset squared difference value threshold, retrieve the absolute timestamps bound to each of the N groups of frames, and subtract the timestamp of the previous frame from the timestamp of the later frame to obtain the time difference values ​​of the N groups of the same waveform, where N is the rated number of frames retrieved within the preset unit time. The three types of parameters included in the generation of time-series delay feature data are: The time difference between adjacent similar waveforms is the set of time difference values ​​corresponding to all successfully matched adjacent highly similar waveforms. Waveform delay frequency is the total number of times the same time difference is repeated. The distribution density of fixed-delay repetitive waveforms is the percentage of frames with fixed delay and repetitive waveforms within the preset unit of time. The three types of parameters are combined into a structured dataset.

3. The method for extracting key information from fire alarm voice messages according to claim 2, characterized in that: The step of marking a waveform region with no fixed time delay and no repeating homologous waveforms as an alarm region, and generating speech integrity timing verification data based on the current waveform region segment, includes: Traverse all frames of the original audio, filter out the audio frames corresponding to the alarm area, and reassemble them according to the order of collection time to form an ordered sequence of human voice frames. The method for determining whether there are gaps in the human voice punctuation within a single frame of the ordered human voice frame sequence includes: Extract the amplitude values ​​of all sampling points within a single frame and compare them with a preset threshold for determining the amplitude of human voice interruption. Determine the sampling interval where the amplitude values ​​of multiple consecutive sampling points within the frame are lower than the interruption threshold as a sound gap within the single frame. Record the frame number corresponding to the gap. Mark the start and end sampling times within the frame as a time-series interruption point group. Arrange all the identified time-series interruption point groups in chronological order to obtain the time-series interruption point distribution data. The effective frame ratio of human voice is calculated statistically, and the total number of complete frames of the original voice and the total number of frames marked as effective alarm personnel voice regions are obtained. The effective frame ratio of human voice is obtained by dividing the total number of frames of the effective alarm personnel voice regions by the total number of complete frames. Calculate the length of the coughing and choking interruption interval corresponding to each of the time-series discontinuity point groups, retrieve the sampling start time stamp and sampling end time stamp, and subtract the sampling start time stamp from the sampling end time stamp to obtain the duration of a single interruption; traverse all the time-series discontinuity points, record the duration of each single interruption, and summarize the entire duration set as the coughing and choking interruption interval length data. The temporal discontinuity point distribution data, the proportion of effective human voice frames, and the length of coughing and choking intervals are respectively output as three rows of sub-data for speech integrity temporal verification data. The three rows of sub-data share the frame timestamp and frame number of the original speech as a unified index.

4. The method for extracting key information from fire alarm voice messages according to claim 3, characterized in that: The step-by-step hierarchical noise reduction processing based on the dual-weighted dynamic adjustment data includes: Based on the timing verification data of the voice integrity, the timing discontinuity point groups within the alarm area are located. Combined with the timing delay feature data, residual echo pseudo-waveforms and pulse interference noise between adjacent timing discontinuity point groups are removed. The human voice timing waveform features and frequency band distribution features between adjacent timing discontinuity point groups are extracted as repair benchmark feature data. Based on the repair benchmark feature data, the timing waveform smoothing and completion processing is performed on the discontinuous voice information chain, and the timing connection calibration processing is performed on the discontinuously split voice information chain.

5. The method for extracting key information from fire alarm voice messages according to claim 1, characterized in that: The process of determining the type of noise interference and marking the region based on the time delay feature data includes: If a waveform with a fixed time delay is detected and repeated waveforms of the same source appear a preset threshold number of times, it is determined to be multiple echo interference in the corridor, and the echo time delay value and repetition density are counted.

6. The method for extracting key information from fire alarm voice messages according to claim 5, characterized in that: The dual-weighted dynamic adjustment data also has graded noise reduction weights corresponding to multiple echoes. After detecting that the frequency of waveform delay and the distribution density of fixed-delay repetitive waveforms are both higher than the threshold based on the time delay feature data, the graded noise reduction weights corresponding to multiple echoes are increased to fix the highest noise reduction priority of multiple echoes. At the same time, the global voice integrity constraint weights are increased simultaneously. The remaining noises are matched with corresponding graded noise reduction weights according to their respective detected time fluctuation intensity. Graded noise reduction processing is performed step by step according to the dual-weighted dynamic adjustment data. Based on the increased graded noise reduction weights of multiple echoes, the delayed and repetitive reflection waveforms are targeted and stripped, while retaining the voice time discontinuity and coughing auxiliary waveform features in the alarm area.

7. The method for extracting key information from fire alarm voice messages according to claim 1, characterized in that: The process involves acquiring the noise-reduced speech signal, executing a mode switching command, and generating real-time feedback time-series data for noise-reduced output, including: By continuously comparing the overall distribution offset of waveform time-series distortion feature parameters of adjacent frames, synchronous abrupt change nodes of multi-dimensional waveform distortion features are captured. Based on the consistency of abrupt changes in the overall waveform distortion distribution, the dynamic switching behavior of the fire scene is determined and a mode switching command is generated.

8. The method for extracting key information from fire alarm voice messages according to claim 1, characterized in that: The step of extracting key fire information based on the target alarm voice includes: Based on the frequency of waveform delay occurrence, the number of echoes corresponding to different propagation directions is distinguished. Based on the number of echoes in different directions, the distance between the alarm person's sound source and the corresponding wall is obtained, and the spatial dimensions of each direction in the corridor are determined. At the same time, based on the time difference of the adjacent similar waveforms in the same vertical direction and the distribution density of the fixed-delay repeated waveforms, the farthest distance from the alarm person's current position to the floor below is obtained.

9. A fire alarm voice key information extraction system, characterized in that, include: The data acquisition module is used to respond to fire alarm call requests, continuously acquire the original voice within the call request, generate original voice time-series waveform data, perform frame-by-frame parsing processing on the original voice time-series waveform data, compare waveform similarity and waveform occurrence time difference frame by frame, and extract time-series delay feature data. The time-series delay feature data includes the time difference value of adjacent similar waveforms, waveform delay occurrence frequency, and distribution density of fixed-delay repeating waveforms. The noise determination module is used to determine the type of noise interference and mark the region based on the timing delay feature data. It marks the waveform region where the detected waveform has no fixed time delay and no repeating homologous waveforms as an alarm region, and generates voice integrity timing verification data based on the current waveform region segment. The weight adjustment module is used to dynamically generate dual-weight dynamic adjustment data based on the timing delay feature data and the speech integrity timing verification data, and to perform hierarchical noise reduction processing step by step according to the dual-weight dynamic adjustment data. The key information extraction module is used to execute mode switching instructions, generate noise reduction output real-time feedback time-series data, compare the noise reduction output real-time feedback time-series data with preset standard alarm voice time-series features, iteratively fine-tune the dual-weight dynamic adjustment data based on the comparison results, and cyclically perform graded noise reduction processing until a target alarm voice that meets the preset recognition information is extracted, and extract key fire information based on the target alarm voice.

10. A storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the method for extracting key information from fire alarm voices as described in any one of claims 1-8.

Images