Method for optimizing design of ssr myoelectric electrode patch based on facial neck vocal muscle group activation

Abstract

The present application belongs to the technical field of silent speech recognition, and specifically relates to an SSR myoelectric electrode patch optimization design method based on face-neck vocal muscle group activation. The present application selects the zygomatic muscle, buccinator muscle, levator anguli oris muscle, platysma muscle, sternocleidomastoid muscle, cricothyroid muscle, thyrohyoid muscle, geniohyoid muscle and depressor labii inferioris muscle closely related to the vocal process as the target muscle group, takes the face, neck and chin area on both sides as the myoelectric signal collection area and pastes a 320-channel high-density surface myoelectric collection array, so as to obtain high-precision vocal muscle group neural discharge information and muscle movement information from the physiological signal level; the active area and the non-active area are determined by quantifying the spatial activation mode of the vocal muscle group, and the electrode layout is optimized by increasing the channel density of the active area and reducing the coverage area of the non-active area; through the electrode optimization layout, the redundant channels are reduced and the calculation efficiency is improved while the high-density electrode array can capture rich muscle activity characteristics.

Description

Technical Field

[0001] This invention belongs to the field of silent speech recognition technology, specifically involving an optimized design method for electromyographic electrode patches based on the activation patterns of facial and neck vocal muscles. Background Technology

[0002] Language is a vital tool for interpersonal communication and emotional expression. Over the past few decades, extensive research has been conducted on automatic speech recognition systems that use acoustic signals as input for speech recognition tasks. However, currently widely used automatic speech recognition systems have several shortcomings. First, because the quality of sound signals is easily affected by environmental noise and non-target speech, the system's recognition performance may drop sharply in certain uncontrollable environments. Second, if sound signals are used in highly confidential settings such as banks and the military, there may be a risk of information leakage. Finally, due to the lack of necessary speech input, patients with aphasia, including those who have undergone laryngectomy, stroke, Parkinson's disease, or locked-in syndrome, cannot use automatic speech recognition systems. Therefore, there is an urgent need to explore silent speech recognition (SSR) technology that does not rely on acoustic signals, generating intelligible non-acoustic signals by acquiring signals related to the vocalization process.

[0003] Surface electromyography (sEMG) is a bioelectrical signal generated during neuromuscular activity and is now widely used in assistive medicine, rehabilitation medicine, and human-computer interaction. Since human language is a direct result of the activation of the vocal muscles, different sEMG signals from the facial and neck muscles during phonation correspond to different spatial activation patterns. Therefore, corresponding speech commands can be identified by collecting and analyzing the electromyographic signals of relevant muscle groups during phonation. Compared with invasive methods such as needle electrode electromyography and electromagnetic articulation imaging, sEMG has advantages such as non-invasive detection and high operability. Compared with electroencephalography (EEG) signals, sEMG has better resistance to physiological fluctuations, higher spatial resolution, and higher recognition accuracy. Therefore, sEMG-based speech recognition (SSR) technology has become a substitute or supplement to traditional automatic speech recognition systems in certain special scenarios.

[0004] In recent years, researchers both domestically and internationally have made many attempts at SSR technology based on sEMG. Solutions similar to this invention mainly include:

[0005] (1) A speech rehabilitation training system for aphasic patients based on facial and neck surface electromyography. The method uses discrete 5-channel electrodes. Five volunteers with speech disorders achieved 89.4% classification of 9 Thai syllables.

[0006] (2) A silent speech decoding method based on surface electromyography of the face and neck, which uses four high-density (HD) flexible electrode arrays placed on the left and right sides of the face and neck to form a 64-channel array.

[0007] The main drawbacks of existing sEMG-based SSR technologies are as follows:

[0008] Defect 1: The articulation process involves complex neuromuscular activities of the articulatory muscle groups, and the related muscle groups are anatomically close to each other. Early SSRs typically used discrete surface electromyography (SEM) electrodes with a limited number of channels to record muscle activity information of the facial and neck articulatory muscles. However, the placement of existing discrete SEM electrodes is usually selected empirically based on physiological anatomy, lacking prior quantitative analysis. Furthermore, due to the limited number of channels, these discrete SEM electrodes cannot cover a sufficiently wide muscle area, resulting in an inability to fully characterize the spatial activation patterns of the articulatory muscles, potentially leading to inaccuracies in speech recognition results.

[0009] Defect 2: Although high-resolution surface electromyography (SSR) electrode arrays can provide high-resolution muscle activation information and improve SSR performance with advancements in electrode fabrication and data acquisition technologies, the redundant channels introduced by HD electrode arrays significantly increase computational complexity, hindering the widespread adoption of SSR technology. Therefore, reducing redundant channels and optimizing electrode placement while maintaining HD performance remains a key research challenge. Summary of the Invention

[0010] To address the shortcomings of the existing technologies, the present invention aims to propose an optimized design method for silent speech recognition electromyographic electrode patches based on the spatial activation pattern of facial and neck vocal muscles, which offers high computational efficiency and good speech recognition performance.

[0011] This invention provides an optimized design method for electromyographic electrode patches for silent speech recognition based on the activation of facial and neck vocal muscles. The method selects the zygomaticus major, buccinator, levator anguli oris, platysma, sternocleidomastoid, cricothyroid, thyrohyoid, mentalis, and depressor labii inferioris muscles, which are closely related to speech, as target muscle groups. A 320-channel HD surface electromyographic array is applied to the facial, lateral neck, and chin areas to acquire electromyographic signals, obtaining high-resolution neural discharge information of the vocal muscles and acquiring richer information on speech-related muscle movement at the physiological signal level. By quantifying the spatial activation patterns of the vocal muscles, active and inactive areas are identified, and the electrode layout is optimized by increasing the channel density in active areas and reducing the coverage area in inactive areas. Through this optimized electrode layout, the method reduces redundant channels and improves computational efficiency while maintaining the ability of the HD electrode array to capture rich muscle activity features. The specific steps are as follows:

[0012] Step 1: Preliminary design of electrode patches for phoneme signal acquisition

[0013] A 320-channel HD electrode array was used to acquire sEMG of each phoneme generated when the user was speaking silently. The preliminary design of the electrode patch layout is as follows: four 8×8 HD electrode arrays (Adhesive Matrix ELSCH064NM1, OTbioelectrononica, Torino, Italy) were placed at the center of both sides of the face and neck, respectively; one 5×13 HD electrode array was placed at the center of the chin, with the upper edge of the electrode array parallel to the lower lip; the two electrodes on the face were labeled A1 and A2, the two electrodes on the neck were labeled A3 and A4, and the electrode at the center of the chin was labeled B1; the electrode placement positions are as follows. Figure 1 As shown in (a), the gel electrode channels of the 8×8 HD electrode array are designed in an elliptical shape with a certain spacing between adjacent electrode channels (e.g., a center-to-center distance of 10 mm); the gel electrode channels of the 5×13 HD electrode array are also designed in an elliptical shape with a certain spacing between adjacent electrode channels (e.g., a center-to-center distance of 2.5 mm). Electrodes A1 and A2 are used to record the activity of the zygomaticus major, buccinator, and levator deltoid muscles. Electrodes A3 and A4 are used to record the activity of the platysma, sternocleidomastoid, cricothyroid, and thyrohyoid muscles. Electrode B1 is used to record the activity of the mentalis and depressor labii inferioris muscles. Additionally, a reference electrode is placed on the mastoid process behind the ear. The HD electrode array is fixed to the skin surface using medical tape and electrode patches.

[0014] Step 2: Phoneme Signal Acquisition

[0015] HD sEMG signals were acquired using the Quattrocento system (OT bioelectrononica, Torino, Italy), which features a 16-bit ADC resolution, a sampling rate of 2048 Hz, and a gain of 150. Subjects pronounced sounds according to a pre-defined guided procedure.

[0016] Step 3: Preprocessing of the acquired phoneme signals

[0017] The starting and ending points of the electromyography (EMG) signal for each silent phonation task are detected based on the trigger signal, thereby segmenting the EMG signal for the silent phonation task and preprocessing it. This includes filtering using high-pass and low-pass filters. The high-pass filter is used to reduce motion artifacts, and the low-pass filter is used to reduce high-frequency noise. Then, a set of notch filters is used to attenuate 50 Hz power line interference and its harmonic components. Continuously stable signals are retained for subsequent extraction of spatial activation pattern features.

[0018] Step 4: Extract spatial activation pattern features of phoneme signals

[0019] The spatial distribution of muscle activity is characterized using the root mean square (RMS) feature of electromyography (EMG) task signals. For each silent vocalization task, five 64-dimensional feature vectors are constructed and normalized to a mean of 0 and a standard deviation of 1. Each feature vector corresponds to one electrode, and each dimension corresponds to a specific channel. The formula for the RMS feature is as follows:

[0020] , (1)

[0021] In the formula, N represents the discrete values ​​of the acquired electromyography (EMG) signals, and N is the data length. The RMS features of each EMG task signal are channel-aligned according to the relative positions of each channel in the HD electrode array, and converted into the spatial distribution of the actual HD electrode array, known as the RMS feature map.

[0022] Step 5:

[0023] Based on RMS feature maps, the spatial activation patterns of the vocal muscles are quantified to identify active and inactive regions. Electrode layout design is optimized by increasing channel density in active regions and reducing coverage area in inactive regions. This electrode layout optimization allows for the reduction of redundant channels and improved computational efficiency while maintaining the ability of HD electrode arrays to capture rich muscle activity features.

[0024] This invention proposes an optimization scheme for the coverage area and channel density of electromyographic electrode arrays based on the spatial activation patterns of facial and neck vocal muscles. By increasing the channel density in active areas and reducing the coverage area in inactive areas, the electrode design is optimized. While maintaining the ability of high-density electrode arrays to capture rich muscle activity features, redundant channels are reduced, computational efficiency is improved, and speech recognition accuracy is enhanced. This makes the electrode patch design for SSR more operable and user-friendly, which is conducive to the development, promotion, and application of SSR systems. Attached Figure Description

[0025] Figure 1 The image shows the application location and design diagram of the HD electrode array. (a) shows the application location of the HD electrode array, and (b) shows the design diagram of the HD electrode array.

[0026] Figure 2 This is a diagram of the experimental guidance procedure.

[0027] Figure 3 The experiment used a randomized phoneme instruction diagram, collecting 14 vowel pronunciation tasks and 15 consonant pronunciation tasks.

[0028] Figure 4The following are the average RMS feature maps of 10 subjects under different phoneme categories. Among them, (a) is the average RMS feature map of 10 subjects under the vowel phoneme category, and (b) is the average RMS feature map of 10 subjects under the consonant phoneme category.

[0029] Figure 5 This is a design diagram of the optimized irregular high-density electrode array. (a) Electrodes A1 and A2, (b) Electrode B1, and (c) Electrodes A3 and A4 are shown. Detailed Implementation

[0030] The present invention proposes a method for designing SSR electromyographic electrode patches based on the spatial activation pattern of facial and neck vocal muscles. The specific steps are as follows:

[0031] Step 1: Electrode layout design for phoneme signal acquisition

[0032] A 320-channel HD electrode array was used to acquire sEMG of the user's silent phonemes. Four 8×8 HD electrode arrays (Adhesive Matrix ELSCH064NM1, OT bioelectrononica, Torino, Italy) were placed at the center of the face and neck, respectively. A 5×13 HD electrode array was placed at the center of the chin, with the upper edge of the array parallel to the lower lip. The two electrodes on the face were labeled A1 and A2, the two electrodes on the neck were labeled A3 and A4, and the electrode at the center of the chin was labeled B1. The electrode placement positions are as follows: Figure 1 As shown in (a), the gel electrode channels of the 8×8 HD electrode array are designed as elliptical (major axis 5 mm, minor axis 2.8 mm), with a center-to-center distance of 10 mm between adjacent electrode channels; the gel electrode channels of the 5×13 HD electrode array are designed as elliptical (major axis 2.2 mm, minor axis 1.5 mm), with a center-to-center distance of 2.5 mm between adjacent electrode channels. The electrode array design diagram is shown below. Figure 1 As shown in (b). Electrodes A1 and A2 were used to record the activity of the zygomaticus major, buccinator, and levator deltoid muscles. Electrodes A3 and A4 were used to record the activity of the platysma, sternocleidomastoid, cricothyroid, and thyrohyoid muscles. Electrode B1 was used to record the activity of the mentalis and depressor labii inferioris muscles. Additionally, a reference electrode was placed on the mastoid process behind the ear. The HD electrode array was secured to the skin surface using medical tape and electrode pads.

[0033] Step 2: Phoneme Signal Acquisition

[0034] Ten healthy participants (nine men and one woman, aged 21-35 years, mean age 25.1 ± 3.8 years) were recruited for data collection. All participants had English as their second native language, good English speaking skills, and no pronunciation or hearing impairments. To reduce impedance between the skin and electrodes, the participants' chin, face, and sides of the neck were carefully cleaned with abrasive gel and alcohol pads. HD-sEMG signals were acquired using a Quattrocento system (OT bioelectrononica, Torino, Italy), which features a 16-bit ADC resolution, a sampling rate of 2048 Hz, and a gain of 150. Before the experiment, all participants simulated the silent pronunciation of each phoneme in their most comfortable way until they could perform the silent pronunciation fluently and naturally.

[0035] During data collection, participants sat in comfortable chairs and followed the experimental instructions on the computer screen. The experimental guidance program was as follows: Figure 2 As shown. The experiment collected 14 vowel pronunciation tasks and 15 consonant pronunciation tasks, as follows. Figure 3 As shown. Participants performed two sets of repeated experiments for each phoneme before moving on to the next phoneme. Within each set, participants performed three repeated task-rest pairs. Each task-rest pair consisted of a 1-second silent speech task and a 1-second rest between tasks. Phoneme instructions were randomly arranged, with one phoneme presented to the participant on the screen at a time. To avoid the impact of muscle fatigue on HD-sEMG, the rest time between sets was 5 seconds. A trigger signal was synchronously generated to a specific channel at the start and end of each task to facilitate later cutting of the electromyographic signals. Each participant performed HD-sEMG signals for 84 vowel tasks (14 vowels × 2 sets of experiments × 3 repetitions) and 90 consonant tasks (15 consonants × 2 sets of experiments × 3 repetitions) (if all tasks were performed correctly). Participants could pause the experiment at any time and then resume data collection from the stopped step. If a participant skipped or performed an incorrect task, the experimental assistant was notified, and the data was discarded from the dataset.

[0036] Figure 3 The experiment used a randomized phoneme instruction diagram and collected 14 vowel pronunciation tasks and 15 consonant pronunciation tasks.

[0037] Step 3: Preprocessing the acquired phoneme signals

[0038] The start and end points of the electromyography (EMG) signals for each silent phonation task are detected based on the trigger signal, thereby segmenting the EMG signals for the silent phonation task and preprocessing them. First, filtering is performed using a 10 Hz high-pass Butterworth filter and a 500 Hz low-pass Butterworth filter (zero-phase digital filters, capable of handling both forward and reverse signals, 8th order in each direction). The high-pass Butterworth filter is used to reduce motion artifacts, and the low-pass Butterworth filter is used to reduce high-frequency noise. Then, a set of notch filters is used to attenuate 50 Hz power frequency interference and its harmonic components. The signal within the 0.25 s reaction time before the start of each task is discarded, retaining the remaining 0.75 s of sustained settling time for extracting spatial activation pattern features.

[0039] Step 4: Extract spatial activation pattern features of phoneme signals

[0040] The root mean square (RMS) feature of electromyography (EMG) task signals was extracted to characterize the spatial distribution of muscle activation patterns. Five 64-dimensional feature vectors were constructed and normalized to a mean of 0 and a standard deviation of 1. Each feature vector corresponds to one electrode, and each dimension corresponds to a specific channel. The RMS feature formula is as follows:

[0041] , (1)

[0042] In the formula, Here, n represents the discrete values ​​of the acquired electromyography (EMG) signals, and n is the data length. The RMS features of each EMG task signal are channel-aligned according to the relative positions of each channel in the HD electrode array, converting them into the spatial distribution of the actual HD electrode array. For example, the 64-dimensional feature vectors of electrodes A1, A2, A3, and A4 are converted into 8*8 RMS feature maps. The 64-dimensional feature vector of electrode B1 is converted into a 5*13 RMS feature map. To reduce the impact of outliers, outlier detection is performed on the RMS feature map of each electrode array in each task. Within a specified electrode array, if its RMS value differs from the RMS mean of its neighboring channels (vertex channels have 3 neighboring channels, and electrode array edge channels have 5 neighboring channels) by more than 3 standard deviations, it is defined as an RMS outlier channel. The mean of the neighboring channels is then used to replace the outlier.

[0043] To visually represent the distribution characteristics of the RMS feature map, this invention selects the centroid coordinates of the RMS feature map to characterize the spatial activation mode of the HD-sEMG. The formula for the centroid coordinates is:

[0044] , (2)

[0045] In the formula, and It is the RMS feature map in and The centroid coordinates in the direction, This represents the element in the i-th row and j-th column of the RMS feature map. and express of and coordinate.

[0046] To objectively quantify the differences in muscle activation areas and assess the synergistic effect of symmetrical articulatory muscle groups, this invention calculates the similarity of symmetrical articulatory muscle groups within the same phoneme category. The similarity can be calculated using the following formula:

[0047] , (3)

[0048] in, and Let represent the elements in the i-th row and j-th column of RMS feature a and RMS feature b, respectively. and They represent The average value of the elements of feature matrix a The mean of the elements of the feature b matrix. Regions with RMS values ​​greater than the average of the RMS feature maps are defined as active regions. Here, RMS feature a and RMS feature b refer to the RMS features of each electromyographic task signal after channel alignment and conversion into the spatial distribution of the actual HD electrode array after calculating the RMS feature values ​​using formula (1). Here, subscripts a and b represent different RMS feature maps of the two sides of the face or neck for the same phoneme category; for example, the 64-dimensional feature vectors of electrodes A1, A2, A3, and A4 are converted into 8*8 RMS feature maps.

[0049] Figure 4 The average RMS feature maps of 10 subjects under different phoneme categories are shown. As can be seen from the figures, regardless of whether it is a vowel or consonant category, the muscle activation areas of the symmetrical articulatory muscle groups always follow a similar pattern. Specifically, the facial muscle activation areas are concentrated in the zygomaticus major and levator anterior deltoid muscles, the neck muscle activation areas are concentrated in the platysma and sternocleidomastoid muscles, and the chin muscle activation areas are concentrated in the mentalis muscle. Using formula (3), the similarity of the symmetrical articulatory muscle groups was calculated, and the results showed significant similarity. In the average RMS feature maps of the 10 subjects under the vowel phoneme category, the similarity of the facial symmetrical articulatory muscle groups was 0.98, and the similarity of the neck symmetrical articulatory muscle groups was 0.97. In the average RMS feature maps of the 10 subjects under the consonant phoneme category, the similarity of the facial symmetrical articulatory muscle groups was 0.88, and the similarity of the neck symmetrical articulatory muscle groups was 0.91. This indicates that the activation patterns of the symmetrical articulatory muscle groups have a certain synergistic effect during speech.

[0050] Step 5:

[0051] based on Figure 4 The average RMS feature map shown illustrates how this invention quantifies the spatial activation patterns of muscle groups, thereby identifying active and inactive regions. Electrode layout design is optimized by increasing channel density in active regions and reducing coverage area in inactive regions. This electrode layout optimization allows for the reduction of redundant channels and improved computational efficiency while maintaining the ability of HD electrode arrays to capture rich muscle activity features. Figure 5 This is the optimized layout design of the irregular HD electrode array. The electrodes are arranged according to the strength of the activation mode, with a minimum spacing of 1mm between electrodes. Each electrode array has 32 channels, for a total of 160 channels. Specifically, the layout of electrodes A1 and A2 (symmetrical) is as follows, counting from the interface end: 3 rows of 6 channels, 2 rows of 5 channels, and 1 row of 4 channels; the layout of electrode B1 is as follows, counting from the interface end: 2 rows of 7 channels and 2 rows of 9 channels; the layout of electrodes A3 and A4 (symmetrical) is as follows, counting from the interface end: 1 row of 8 channels, 1 row of 7 channels (equivalent to removing the 5th channel from the left in the above 8 channels), 5 rows of 3 channels, and 1 row of 2 channels.

[0052] This invention proposes an optimization scheme for the coverage area and electrode density of SSR electrode array based on the spatial activation mode of facial and neck vocal muscles. By increasing the channel density in active areas and reducing the coverage area in inactive areas, the electrode design is optimized. While maintaining the ability of HD electrode arrays to capture rich muscle activity features, redundant channels are reduced and computational efficiency is improved. This makes the SSR-based electrode patch design more operable and user-friendly, which is conducive to the development, promotion and application of SSR systems.

Claims

1. A method for optimizing the design of SSR electrode patches based on activation of facial and neck vocal muscles, characterized in that, The target muscle groups were selected from the zygomaticus major, buccinator, levator anguli oris, platysma, sternocleidomastoid, cricothyroid, thyrohyoid, mentalis, and depressor labii inferioris muscles, which are closely related to speech. The facial, lateral neck, and chin areas were selected as the acquisition areas for electromyography (EMG) signals, and a 320-channel high-density surface EMG array was applied. This allowed for the acquisition of high-precision nerve discharge and muscle movement information of the speech muscles at the physiological signal level. Active and inactive areas were determined by quantifying the spatial activation patterns of the muscle groups, and the electrode layout was optimized by increasing the channel density in active areas and reducing the coverage area in inactive areas. Through electrode optimization, the rich muscle activity characteristics captured by the high-density electrode array were captured while reducing redundant channels and improving computational efficiency. The specific steps are as follows: Step 1: Preliminary design of the electrode patch for acquiring phoneme signals A 320-channel high-density electrode array was used to acquire sEMG signals generated when the user was silently uttering phonemes. The preliminary layout of the electrode patches was designed as follows: four 8×8 high-density electrode arrays were placed at the center of the face and neck, respectively; one 5×13 high-density electrode array was placed at the center of the chin, with the upper edge of the electrode array parallel to the lower lip. The two electrodes on the face were designated A1 and A2, the two electrodes on the neck were designated A3 and A4, and the electrode on the chin was designated B1. The gel electrode channels of the 8×8 high-density electrode array were designed to be elliptical, with a certain spacing between adjacent channels. The gel electrode channels of the 5×13 high-density electrode array were also designed to be elliptical, with a certain spacing between adjacent channels. Electrodes A1 and A2 were used to record muscle activity signals of the zygomaticus major, buccinator, and levator deltoid muscles. Electrodes A3 and A4 were used to record muscle activity signals of the platysma, sternocleidomastoid, cricothyroid, and thyrohyoid muscles. Electrode B1 was used to record muscle activity signals of the mentalis and depressor labii inferioris muscles. In addition, a reference electrode was placed on the mastoid process behind the ear. Step 2: Phoneme Signal Acquisition High-density sEMG signals were acquired using the Quattrocento system, and the subjects made sounds according to a specific experimental procedure. Step 3: Preprocess the acquired phoneme signals The starting and ending points of the electromyography (EMG) signal for each silent phonation task are detected based on the trigger signal, thereby segmenting the EMG signal of the silent phonation task and preprocessing it, including filtering with a high-pass filter and a low-pass filter; the high-pass filter is used to reduce motion artifacts, and the low-pass filter is used to reduce high-frequency noise; then, a set of notch filters is used to attenuate the 50 Hz power frequency interference and its harmonic components, retaining a continuous and stable signal for subsequent extraction of spatial activation mode features. Step 4: Spatial activation mode feature extraction of phoneme signals The root mean square (RMS) feature of electromyography (EMG) task signals is used to characterize the spatial distribution of muscle activation patterns. For each silent vocalization task, five 64-dimensional feature vectors are constructed and normalized to a mean of 0 and a standard deviation of 1. Each feature vector corresponds to one electrode, and each dimension corresponds to a specific channel. The formula for the RMS feature is as follows: ， （1） In the formula, The discrete values ​​of the acquired electromyography (EMG) signals are represented by n, which is the data length. The RMS features of each EMG task signal are channel-aligned according to the relative positions of each channel in the high-density electrode array, and converted into the spatial distribution of the actual high-density electrode array, which is called the RMS feature map. Step 5: Based on RMS feature maps, the spatial activation patterns of the vocal muscles are quantified to identify active and inactive regions. The electrode layout design is optimized by increasing the channel density of active regions and reducing the coverage area of ​​inactive regions. Through electrode layout optimization, redundant channels are reduced and computational efficiency is improved while maintaining the ability of high-density electrode arrays to capture rich muscle activity features.

2. The SSR electrode patch optimization design method based on facial and neck vocal muscle activation according to claim 1, characterized in that, Step four involves converting the RMS feature map of the actual high-density electrode array into a spatial distribution map. Specifically, the 64-dimensional feature vectors of electrodes A1, A2, A3, and A4 are converted into 8*8 RMS feature maps, and the 64-dimensional feature vector of electrode B1 is converted into a 5*13 RMS feature map. To reduce the impact of outliers, outlier detection is performed on the RMS feature map of each electrode array in each task. Within a specified electrode array, if the RMS feature value of a certain channel differs from the average RMS feature value of its adjacent channels by more than 3 standard deviations, it is defined as an RMS outlier. Then, the average RMS feature value of its adjacent channels is used to replace the outlier.

3. The SSR electrode patch optimization design method based on facial and neck vocal muscle activation according to claim 2, characterized in that, To visually represent the distribution characteristics of RMS feature maps, the centroid coordinates of the RMS feature maps are selected to characterize the spatial activation patterns of high-density surface electromyography signals; the formula for the centroid coordinates is: ， （2） In the formula, and It is the RMS feature map in and The centroid coordinates in the direction, This represents the element in the i-th row and j-th column of the RMS feature map. and express of and coordinate; To objectively quantify the differences in muscle activation areas and assess the synergistic effect of symmetrical articulatory muscle groups, the similarity of symmetrical articulatory muscle groups within the same phoneme category was calculated: ， （3） in, and Let represent the elements in the i-th row and j-th column of RMS feature a and RMS feature b, respectively. and They represent The mean and sum of the elements of the feature matrix a The mean of the elements of the feature matrix b; the RMS value is greater than... The region where the feature map average is averaged is defined as the active region; here, subscripts a and b represent different RMS feature maps on the two sides of the face or neck for the same phoneme category, respectively. Will Defined as the average RMS feature map of all correctly performed vowel pronunciation tasks by 10 subjects. Defined as the average RMS feature map of all correctly performed silent consonant articulation tasks by 10 subjects; each Figures and The figures were normalized to between 0 and 1. Regardless of whether they were vowels or consonants, the muscle activation areas of the symmetrical articulatory muscle groups always followed a similar pattern. Specifically, the muscle activation areas of the face were concentrated in the zygomaticus major and levator anguli oris muscles, the muscle activation areas of the neck were concentrated in the platysma and sternocleidomastoid muscles, and the muscle activation areas of the chin were concentrated in the mentalis muscle. Using formula (3), the similarity of the symmetrical articulatory muscle groups was calculated, and the results showed significant similarity. This indicates that the activation patterns of the symmetrical articulatory muscle groups have a synergistic effect during speech.

4. The SSR electrode patch optimization design method based on facial and neck vocal muscle activation according to claim 2, characterized in that, In step five, the electrodes are arranged according to the strength of the activation mode, with a spacing of no less than 1 mm between the electrodes. Each electrode array has 32 channels, for a total of 160 channels. Among them, electrodes A1 and A2 are symmetrical, with the following layout from the interface end: 3 rows of 6 channels, 2 rows of 5 channels, and 1 row of 4 channels. Electrode B1 is arranged with the following layout from the interface end: 2 rows of 7 channels and 2 rows of 9 channels. Electrodes A3 and A4 are symmetrical, with the following layout from the interface end: 1 row of 8 channels, 1 row of 7 channels, 5 rows of 3 channels, and 1 row of 2 channels.

Images