Acoustic device

Abstract

The present invention relates to an acoustic device for sound direction and content recognition, comprising: a body comprising a gap defined by an inner boundary of the body; a cantilever extending into the gap from the inner boundary and comprising a free end in the gap; at least one pair of inner electrodes arranged at the free end of the cantilever; at least one pair of outer electrodes arranged at the inner boundary of the body; and a plurality of piezoelectric nanofibers extending radially from the free end of the cantilever to the inner boundary and arranged asymmetrically around the free end, wherein the piezoelectric nanofibers are configured to convert an acoustic signal into an electrical signal detectable using the at least one pair of inner electrodes and / or the at least one pair of outer electrodes.

Description

[0001] Acoustic Device

[0002] Technical Field

[0003] The present invention is concerned with an acoustic device, specifically an acoustic device suitable for sound direction and content recognition.

[0004] Sound recognition systems are one of the most promising human-machine interfaces due to its unique advantages in contactless, personalized and bilateral communications, which have attracted many interests in internet of things (loT), body area networks (BAN) and smart home appliances. Sound recognition systems conventionally involve acoustic hardware in the form of an acoustic sensor, and recognition software functionalizing as harvesting, recognizing and interacting of sound waves based on the contents or directions.

[0005] Existing acoustic sensors are commonly based on condensers or microphones. Such devices can be effective in detecting sound. However, acoustic devices based on condensers or microphones usually present relatively low biocompatibility, low sensitivity and high power consumption. Moreover, such devices do not typically provide for accurate sound direction and / or content recognition.

[0006] Modern acoustic sensors may also be based on piezoelectric nanofibers, which transduce sound waves to electrical signals. Such devices may demonstrate heightened sensitivity compared with conventional acoustic sensors based on condensers or microphones. However, existing piezoelectric acoustic devices still frequently result in inaccurate sound recognition. Moreover, such devices are typically not equipped for accurate sound direction recognition.

[0007] Therefore, there are improvements that can be made in the field of acoustic devices / sensors.of the Invention

[0008] Aspects of the invention are set out in the accompanying claims.

[0009] In accordance with some embodiments described herein, there is provided an acoustic device for sound direction and content recognition, comprising: a body comprising a gap defined by an inner boundary of the body; a cantilever extending into the gap from the inner boundary and comprising a free end in the gap; at least one pair of inner electrodes arranged at the free end of the cantilever; at least one pair of outer electrodes arranged at the inner boundary of the body; and a plurality of piezoelectric nanofibers extending radially from the free end of the cantilever to the inner boundary and arranged asymmetrically around the free end, wherein the piezoelectric nanofibers are configured to convert an acoustic signal into an electrical signal detectable using the at least one pair of inner electrodes and / or the at least one pair of outer electrodes.

[0010] The present arrangement uses a plurality of asymmetrically arranged piezoelectric nanofibers to convert an acoustic signal to an electrical signal for sound content and / or direction recognition, in use. The particular asymmetric arrangement of the piezoelectric nanofibers around a central cantilever generates an electrical signal from which sound direction and / or content can be determined with greater accuracy compared with existing devices. In particular, the asymmetric geometry of the piezoelectric nanofibers around the free end of the cantilever confers directionality and non-linear multiresonant responses of the nanofibers. This permits determination of the contents and direction of sound stimulation with high accuracy, based on the generated electrical signal. Moreover, the asymmetric geometry filters directional spectral features of the acoustic signals, enabling precise sound direction recognition.

[0011] A highly sensitive acoustic device is therefore provided, that can convert and transmit sound into complex nonlinear multi-resonant spectra. This permits sound direction and content recognition with high sensitivity and accuracy. Indeed, the sound spatial recognition of the present device surpasses human hearing capabilities, accurately recognising all sound directions both horizontally and vertically

[0012] In examples, the electrical signal comprises features associated with a direction and content of the acoustic signal.

[0013] The present arrangement further improves the accuracy of sound direction and / or content recognition. The acoustic device is capable of converting and transmitting acoustic signals into electrical signals comprising complex nonlinear multi-resonant spectra varying with the location and content of the acoustic signals. The features associated with the direction and content of the acoustic signal correspond to spectral characteristics of the measured outputvoltage across the electrodes. Therefore, by features associated with a direction and content of the acoustic signal, it is meant spectral features of a measured voltage output across the pair of inner electrodes and / or the pair of outer electrodes. These spectral features may include one or more of frequency components, phase differences, amplitude variations, and harmonic structures that encode spatial information about the incoming acoustic signal. In the present example, the electrical signal generated by the acoustic device therefore contains detailed information regarding the location and / or content of the acoustic signals. This detailed information can be analysed to accurately determine the acoustic signal’s origin and its acoustic characteristics.

[0014] In examples, the plurality of piezoelectric nanofibers comprises a plurality of groups of piezoelectric nanofibers; the length of each piezoelectric nanofiber in a respective group is the same; and the length of the piezoelectric nanofibers in each group is different from the length of the piezoelectric nanofibers in each other group.

[0015] According to the present example, the asymmetric arrangement of the piezoelectric nanofibers is provided by varying the length of the nanofibers around the cantilever. Providing distinct groups of piezoelectric nanofibers with differing lengths further improves the accuracy of sound content and / or direction recognition. This is because the response of each group of piezoelectric nanofibers to an acoustic signal will be expected to be different, thus providing for a more granular electrical signal and thus enhanced readout accuracy.

[0016] In examples, each group of piezoelectric nanofibers is tuned to a resonant frequency of the respective group of piezoelectric nanofibers.

[0017] Tuning the piezoelectric nanofibers to a resonant frequency assists in maximising output of the electrical signal from the device. Greater electrical signal output from the device improves the efficiency and accuracy with which sound content and / or directionality can be determined. In examples, each group of piezoelectric nanofibers is tuned to a different resonant frequency. The present arrangement further improves the accuracy of sound content and / or direction recognition. A piezoelectric acoustic device tuned only to first-order resonant frequency results in non-optimal sound direction determination, particularly for vertical sound localization. By tuning each group of piezoelectric nanofibers to a different resonant frequency (e.g. multiorder frequencies of the acoustic device), greater sound direction discrimination can be achieved based on the generated electrical signal.

[0018] In examples, the acoustic device comprises a plurality of outer electrode pairs arranged at the inner boundary of the body, and each outer electrode pair of the plurality of outer electrode pairs is arranged at a periphery of a respective group of piezoelectric nanofibers.According to the present arrangement, each group of piezoelectric nanofibers has a pair of electrodes arranged at its periphery, specifically at the inner boundary of the body of the acoustic device. This means that each piezoelectric nanofiber group has a dedicated outer electrode pair that can be used to detect the electrical signal generated by a respective group of nanofibers. Thus, a plurality of electrical signals may be generated (one by each group of piezoelectric nanofibers) and detectable using a corresponding pair of outer electrodes, and / or pair of inner electrodes. This further improves the accuracy of the sound content and / or direction recognition, as there is a one-to-one correspondence between each group of piezoelectric nanofibers and the outer electrode pairs.

[0019] In examples, the length of the plurality of piezoelectric nanofibers decreases in a circumferential direction of the free end of the cantilever such that the plurality of piezoelectric nanofibers form a spiral-like shape.

[0020] The present inventors have determined that the spiral-like shape of the plurality of piezoelectric nanofibers enhances the sound direction and / or content recognition accuracy. In particular, the spiral-like shape of the piezoelectric nanofibers has been found to result in improved sound direction and / or content recognition accuracy over piezoelectric nanofibers which are arranged in a random geometry, or piezoelectric nanofibers which are arranged symmetrically.

[0021] In examples, the length of the plurality of piezoelectric nanofibers decreases stepwise in the circumferential direction.

[0022] Decreasing the length of the plurality of piezoelectric nanofibers decreases stepwise provides for groups of nanofibers having the same length whilst retaining the overall spiral-like shape of the plurality of piezoelectric nanofibers. This provides for simplified signal data collection and analyses, while achieving high accuracy sound content and / or direction recognition. In examples, the length of the plurality of piezoelectric nanofibers decreases continuously in the circumferential direction.

[0023] The present arrangement provides for a granular output from the acoustic device, as length (and thus expected response) of the piezoelectric nanofibers changes continuously. Therefore, high sound frequency selectivity can be achieved, and sound direction and / or content recognition accuracy can be improved.

[0024] In examples, a number of pairs of the outer electrodes is equal to a number of the piezoelectric nanofibers, and each piezoelectric nanofiber of the plurality of piezoelectric nanofibers extends to a respective outer electrode pair.The present arrangement provides for a very granular electrical signal generated by the acoustic device, as there is a one-to-one correspondence between each piezoelectric nanofiber and outer electrode pair. Therefore, sound direction and / or content can be determined with high accuracy based on such an electrical signal.

[0025] In examples, the body further comprises a polymer layer between each electrode of the at least one pair of inner electrodes, and between each electrode of the at least one pair of outer electrodes.

[0026] The polymer layer boosts the electrical signal output by the acoustic device. Improved electrical signal output improves the efficiency and accuracy with which a sound direction and / or content can be determined.

[0027] In examples, the acoustic device is a self-powered acoustic device.

[0028] The present example provides a device that can be continually operated, and does not require frequent replacement of a battery, for example. This improves the versatility of the device, as it can be deployed in areas where traditional power solutions are impractical. It also improves the reliability of the device, as it does not rely on an external power source to operate.

[0029] In examples, the plurality of piezoelectric nanofibers are piezoelectric polymer nanofibers comprising piezoelectric nanoparticles.

[0030] The piezoelectric nanofibers according to the present example are particularly robust. Therefore, the reliability and durability of the acoustic device is improved.

[0031] In examples, the piezoelectric polymer is poly(vinylidene fluoride-co-trifluoroethylene) copolymer (PVDF-TrFE) or polyvinylidene fluoride (PVDF); the piezoelectric nanoparticles comprise barium titanate (BTO) nanoparticles; and a concentration of the BTO nanoparticles is 0-12 wt% relative to the weight of the PVDF-TrFE or PVDF.

[0032] The addition of BTO nanoparticles to the polymer result in an increase in the apparent modulus of the nanofibers, however decrease the strengths and strain of the nanofibers due to the interfaces between the polymer and the BTO nanoparticles. The present arrangement balances the apparent modulus, strength and strain of the nanofibers by providing the BTO nanoparticles at a preferred concentration of 0-12 wt% relative to the weight of the PVDF-TrFE or PVDF.

[0033] In a further embodiment, a system comprising the acoustic device according to any of the above examples is provided. The system is configured to communicate with one or more data processors configured to: receive the electrical signal from the acoustic device; and processthe electrical signal using a machine learning model to determine at least one of a direction and content of the acoustic signal based on features of the electrical signal.

[0034] According to the present arrangement, a complete system is provided that is capable of determining a sound direction and / or content from the electrical signal generated by the acoustic device. In particular, the electrical signal is encoded by a machine learning model (e.g. a neural network) to enable sound direction and / or content recognition based on the contents of the electrical signal.

[0035] In examples, the one or more data processors are configured to: extract features associated with at least one of the direction and content of the acoustic signal from the electrical signal; input the extracted features into the machine learning model trained for at least one of sound direction and content classification and regression; and determine at least one of the direction and content of the acoustic signal based on an output from the machine learning model. According to the present arrangement, sound direction and / or content recognition is achieved in an efficient and accurate manner using a trained machine learning model. The machine learning model is capable of learning and identifying the specific spectral features of voltage signals collected as the location of the sound source is changed. The model is used to capture the essential nonlinearities of the acoustic features transduced by the vibration of the asymmetric piezoelectric nanofibers, resulting in accurate sound direction recognition and / or sound content recognition.

[0036] In examples, the one or more data processors are configured to pre-process the electrical signal prior to processing the electrical signal using the machine learning model; and to pre-process the electrical signal, the one or more data processors are configured to at least one of: detrend the electrical signal for input to the machine learning model; split the electrical signal for input to the machine learning model; and apply a short-time Fourier transform (STFT) function to the electrical signal to generate a plurality of STFT signals and / or merge the STFT signals for input to the machine learning model.

[0037] According to the present arrangement, the electrical signal generated by the acoustic device is pre-processed prior to input to the machine learning model. This improves the quality and reliability of the signal which is input to the machine learning model, thus permitting the model to generate an output which is representative of a sound direction and / or content. Thus, the sound direction and / or content is determined with improved accuracy and efficiency.

[0038] In examples, the machine learning model is trained based on historical electrical signals generated by the acoustic device. A highly accurate machine learning model is therefore provided.In a further embodiment, a method of determining at least one of a direction and content of an acoustic signal is provided. The method comprises steps of: receiving, by an acoustic device comprising a plurality of asymmetrically arranged piezoelectric nanofibers, an acoustic signal; converting, by the acoustic device and using the plurality of asymmetrically arranged piezoelectric nanofibers, the acoustic signal to an electrical signal comprising features associated with at least one of a direction and content of the acoustic signal; outputting, by the acoustic device and to one or more data processors, the electrical signal; receiving, by one or more data processors, the electrical signal; and processing, by the one or more data processors, the electrical signal using a machine learning model to determine at least one of the direction and content of the acoustic signal based on the features of the electrical signal. For at least the same reasons discussed above with respect the example acoustic devices and systems, the present method provides for accurate sound direction and / or content recognition.

[0039] In examples, the acoustic device used in the above-discussed method is the acoustic device according to any of the example acoustic devices discussed above.

[0040] In examples, the processing comprises step of: extracting features associated with at least one of the direction and content of the acoustic signal from the electrical signal; inputting the extracted features into the machine learning model trained for at least one of sound direction and content classification and regression; and determining at least one of the direction and content of the acoustic signal based on an output from the machine learning model.

[0041] In examples, the method further comprises pre-processing, by the one or more data processors, the electrical signal prior to processing the electrical signal using the machine learning model, wherein the pre-processing comprises at least one of: detrending the electrical signal for input to the machine learning model; splitting the electrical signal for input to the machine learning model; and applying a Short-time Fourier transform, STFT, function to the electrical signal to generate a plurality of STFT signals and / or merge the STFT signals for input to the machine learning model.

[0042] In examples, the machine learning model in the above discussed methods is trained based on historical electrical signals generated by the acoustic device.

[0043] In a further embodiment, a method of manufacturing the acoustic device according to any of the above-discussed examples is provided. The method comprises: attaching the at least one pair of inner electrodes onto a lower layer and an upper layer of the body, the upper layer and lower layer comprising the gap; attaching the at least one pair of outer electrodes to the lowerlayer and the upper layer; electrospinning the plurality of piezoelectric nanofibers into the gap of the lower layer; and joining the upper layer and bottom layer together.

[0044] The present approach provides for simple manufacture whilst allowing for the geometry and alignment of the piezoelectric nanofibers to be closely controlled via electrospinning.

[0045] In examples, the method further comprises applying the upper layer of the body and the lower layer of the body on opposing sides of a polymer layer.Brief of the

[0046] One or more embodiments of the invention will now be described, by way of example only, and with reference to the following figures in which:

[0047] Figure 1 shows a perspective view of an acoustic device according to an example of the present disclosure;

[0048] Figure 2 shows an exploded view of an acoustic device according to an example of the present disclosure;

[0049] Figure 3 shows a portion of an acoustic device according to an example of the present disclosure;

[0050] Figure 4 shows a system according to an example of the present disclosure;

[0051] Figure 5 schematically shows a workflow of machine learning based sound direction and content recognition according to an example of the present disclosure;

[0052] Figure 6 schematically shows the structure of machine learning prediction and regression based on an auditory spectrum encoder and sound feature extractor according to an example of the present disclosure;

[0053] Figure 7 shows a method of determining at least one of a direction and content of an acoustic signal according to an example of the present disclosure;

[0054] Figures 8a-e shows plots of experimental data obtained using an acoustic device according to an example of the present disclosure; and

[0055] Figures 9a-d show plots of experimental data obtained using an acoustic device according to an example of the present disclosure.

[0056] Any reference to prior art documents in this specification is not to be considered an admission that such prior art is widely known or forms part of the common general knowledge in the field. As used in this specification, the words “comprises”, “comprising”, and similar words, are not to be interpreted in an exclusive or exhaustive sense. In other words, they are intended to mean “including, but not limited to”. The invention is further described with reference to the following examples. It will be appreciated that the invention as claimed is not intended to be limited in any way by these examples. It will also be recognised that the invention covers not only individual embodiments but also combination of the embodiments described herein.

[0057] The various embodiments described herein are presented only to assist in understanding and teaching the claimed features. These embodiments are provided as a representative sample of embodiments only, and are not exhaustive and / or exclusive. It is to be understood thatadvantages, embodiments, examples, functions, features, structures, and / or other aspects described herein are not to be considered limitations on the scope of the invention as defined by the claims or limitations on equivalents to the claims, and that other embodiments may be utilised and modifications may be made without departing from the spirit and scope of the claimed invention. Various embodiments of the invention may suitably comprise, consist of, or consist essentially of, appropriate combinations of the disclosed elements, components, features, parts, steps, means, etc, other than those specifically described herein. In addition, this disclosure may include other inventions not presently claimed, but which may be claimed in future.Detailed

[0058] The present disclosure provides a piezoelectric acoustic device configured to convert an acoustic signal to an electrical signal, from which a sound direction and / or content can be determined with high accuracy.

[0059] Referring to Figure 1 , there is shown a perspective view of an acoustic device 2 according to an example of the present disclosure. The acoustic device 2 is particularly suitable for use in sound direction and content recognition.

[0060] The acoustic device 2 comprises a body 4. The body 4 comprises a gap 6 therein. The gap 6 is defined by an inner boundary 8 of the body 4. The body 4 may be formed from a polymer material, such as polyimide or the like.

[0061] The acoustic device 2 further comprises a cantilever 10. The cantilever 10 extends into the gap 6 from the inner boundary 8. The cantilever 10 comprises a free end 12 in the gap 6.

[0062] The acoustic device 2 further comprises at least one pair of inner electrodes 14 arranged at the free end 12 of the cantilever 10. The inner electrodes 14 are arranged vertically opposite each other within the free end 12 of the cantilever 10, as shown in Figure 2 by way of electrodes 14a and 14b forming the pair of inner electrodes 14. That is, the pair of inner electrodes 14 are embedded within the free end 12 of the cantilever 10.

[0063] The acoustic device 2 further comprises at least one pair of outer electrodes 16 arranged at the inner boundary 8 of the body 4. The outer electrodes 16 are arranged vertically opposite each other within the body 4 at the inner boundary 8, as shown in Figure 2 by way of electrodes 16a and 16b forming the pair of outer electrodes 16. That is, the at least one pair of outer electrodes 16 are embedded within the body 4 at the inner boundary 8.

[0064] The acoustic device 2 further comprises a plurality of piezoelectric nanofibers 18. The plurality of piezoelectric nanofibers 18 extend radially from the free end 12 of the cantilever 10 to the inner boundary 8 of the body 4. The plurality of piezoelectric nanofibers 18 are arranged asymmetrically around the free end 12 of the cantilever 10.The plurality of piezoelectric nanofibers 18 are configured to convert an acoustic signal into an electrical signal. The electrical signal is detectable using the at least one pair of inner electrodes 14 and / or the at least one pair of outer electrodes 16.

[0065] In use, sound received at the acoustic device 2 causes the piezoelectric nanofibers 18 to vibrate. Upon acoustic vibrations, electric dipoles are induced in the piezoelectric nanofibers 18. The electric potential is then built up along each nanofiber and measured using the inner pair of electrodes 14, the outer pair of electrodes 16, or both. The electric potential (i.e. the electrical signal) carries both rich information of fibre vibration and sound signal stimulation, which can be analysed to determine a content and / or direction of the sound.

[0066] In particular, the asymmetric arrangement of the piezoelectric nanofibers 18 around the free end 12 of the cantilever enables the acoustic device 2 to convert acoustic signals into electrical signals comprising complex nonlinear multi-resonant spectra varying with the location and content of the acoustic signals. This permits accurate sound content and / or direction determination using the device 2.

[0067] The features associated with the direction and content of the acoustic signal correspond to spectral characteristics of the measured voltage output across the inner pair of electrodes 14 (comprising electrodes 14a and 14b) and / or the outer pair of electrodes 16 (comprising electrodes 16a and 16b). The spectral features may include frequency components, phase differences, amplitude variations, and harmonic structures that encode spatial information about the incoming acoustic signal. Specifically, directional information can be derived from Time Difference (phase shift between signals at different electrodes) and Level Difference (amplitude variation across electrodes), which may be used for localizing the sound source. For speech signals, features such as Formants, Fundamental Frequency, and Prosody can provide further cues to enhance directional recognition and signal classification. By analysing these spectral and temporal properties of the electrical signal generated by the acoustic device 2, it is possible to accurately determine the acoustic signal’s origin and its acoustic characteristics.

[0068] Referring now to Figure 2, there is shown an exploded view of the acoustic device illustrated in Figure 1. Like reference numerals are used in Figures 1 and 2 to denote corresponding elements. As shown in Figure 2, the ends of the piezoelectric nanofibers 18 are positioned between the pair of inner electrodes 14a, 14b, and the pair of outer electrodes 16a, 16b. This permits measurement of the voltage generated by the piezoelectric nanofibers 18 in response to an acoustic signal. Specifically, the voltage across electrodes 14a, 14b can be measured.Additionally or alternatively, the voltage across electrodes 16a and 16b can be measured. The measured voltage I electrical signal comprises spectral features representative of a content and / or direction of the acoustic signal. These spectral features can be analysed by one or more data processors to determine a direction and / or content of the acoustic signal.

[0069] As shown in Figure 2, the acoustic device optionally comprises a polymer layer 20 between each electrode 14a, 14b of the at least one pair of inner electrodes 14, and between each electrode 16a, 16b of the at least one pair of outer electrodes 16. The polymer layer 20 boosts the output signal measured using the pair of inner electrodes 14a, 14b and / or the pair of outer electrodes 16a, 16b. The polymer layer 20 therefore enhances the electrical signal output of the acoustic device 2. This improvement is partly attributed to the small gap between the polymer and the electrodes, which may induce a triboelectric effect. The interaction at this interface increases charge generation and contributes to a stronger signal output.

[0070] The polymer layer 20 may comprise a thermoplastic, such as polyethylene terephthalate (PET) or the like. In other examples, the acoustic device 2 does not comprise the polymer layer 20.

[0071] The acoustic device 2 is formed by attaching the at least one pair of inner electrodes 14a, 14b onto a lower layer 4a and an upper layer 4b of the body 4, respectively. Both the lower layer 4a and the upper layer 4b comprise a gap, which when combined form the gap 8 shown in Figure 1. Specifically, inner electrode 14a is formed on an upper surface of the lower layer 4a at the free end 12 of the cantilever 10, and inner electrode 14b is formed on a lower surface of the upper layer 4b at the free end 12 of the cantilever 10.

[0072] The acoustic device 2 is further formed by attaching the at least one pair of outer electrodes 16a, 16b to the lower layer 4a and the upper layer 4b, respectively. Specifically, outer electrode 16a is formed on an upper surface of the lower layer 4a at or near the boundary 8 of the gap. Outer electrode 16b is formed on a lower surface of the upper layer 4b at or near the boundary 8.

[0073] The electrodes 14, 16 may be attached to the body 4 by heat pressing. The electrodes 14, 16 may be made from copper, for example, or another electrically conductive material.

[0074] The acoustic device 2 is further formed by electrospinning the plurality of piezoelectric nanofibers 18 into the gap of the lower layer 4b. Electrospinng advantageously allows the geometry of the piezoelectric nanofibers 18 to be closely controlled, thus providing for complexasymmetric geometries of the piezoelectric nanofibers 18 around the free end 12 of the cantilever 10.

[0075] Finally, the acoustic device 2 is formed by joining the upper layer 4b and the lower layer 4a together, for example by applying heat and pressure to the layers. This method of manufacturing the device results in the acoustic device 2 illustrated in Figure 1. In examples where polymer layer 20 is present, the upper layer 4b and lower layer 4a are applied on opposing sides of the polymer layer 20, such that the polymer layer 20 is positioned between the upper layer 4b and lower layer 4a.

[0076] Referring now to Figure 3, there is shown a portion of an acoustic device according to an example of the pressure disclosure. Like reference numerals are used in Figure 3 to refer to like elements shown in Figures 1 and 2.

[0077] As shown in the example of Figure 3, the plurality of piezoelectric nanofibers 18 comprise a plurality of groups of piezoelectric nanofibers 18a, 18b, 18c, 18d. Although four groups of piezoelectric nanofibers are illustrated in Figure 3, it will be appreciated that a different number of groups may be implemented such as two, three, or more than four groups.

[0078] As shown in Figure 3, the length of each piezoelectric nanofiber in a respective group is the same. For example, the length of each piezoelectric nanofiber in group 18a is the same. The length of each piezoelectric nanofiber in group 18b is the same, etc. Furthermore, the length the piezoelectric nanofibers in each group is different from the length of the piezoelectric nanofibers in each other group. For example, the length of the piezoelectric nanofibers in group 18a is different from the length of the piezoelectric nanofibers in groups 18b, 18c, and 18d. Likewise, the length of the piezoelectric nanofibers in group 18b is different from the length of the piezoelectric nanofibers in groups 18a, 18c, and 18d. In this way, the asymmetric arrangement of the piezoelectric nanofibers around the free end 12 of the cantilever 10 may be provided by varying the lengths of the piezoelectric nanofibers. In other examples, the asymmetric geometry may be provided by varying the directions of the nanofibers 18, for example. Additionally or alternatively, the diameter and / or composition of the piezoelectric nanofibers 18 can be varied in the device 2 to provide the asymmetric arrangement.

[0079] As an example, the length of the piezoelectric nanofibers in group 18a may be approximately 15 mm. The length of the piezoelectric nanofibers in group 18b may be approximately 20 mm. The length of the piezoelectric nanofibers in group 18c may be approximately 25 mm. The length of the piezoelectric nanofibers in group 18d may be approximately 30 mm. In general, the length of the piezoelectric nanofibers 18 may be on the scale of millimetres.Each group of piezoelectric nanofibers 18a-d may be tuned to a resonant frequency of the respective group of piezoelectric nanofibers. For example, the group 18a may exhibit a resonant frequency of 655 Hz, while groups 18b, 18c and 18d may resonate at frequencies of 310 Hz, 210 Hz, and 190 Hz, respectively. Accordingly, the resonant behaviour of each group of piezoelectric nanofibers may be distinct in the device 2, such that each group of piezoelectric nanofibers 18a-d is tuned to a different resonant frequency.

[0080] The resonant frequency of each group 18a-d can be controlled by varying the lengths and / or radii of the piezoelectric nanofibers. Specifically, vibrational peaks may be observed at relatively low frequencies in groups 18c and 18d, which have relatively larger radii and / or longer lengths. The resonant peaks shift to higher frequencies as the lengths decrease in groups 18a and 18b.

[0081] In the example shown in Figure 3, the acoustic device comprises a plurality of outer electrode pairs 22, 24, 26 and 28 arranged at the inner boundary 8 of the body. For ease of illustration, only the bottom electrode in each electrode pair 22, 24, 26 and 28 is shown in Figure 3. However, corresponding upper electrodes on the upper layer 4b of the body 4 would be present (see e.g. Figure 2, which shows the upper electrodes 16b and lower electrodes 16a).

[0082] Each outer electrode pair of the plurality of outer electrode pairs 22, 24, 26 and 28 is arranged at a periphery of a respective group of piezoelectric nanofibers 18a-d. For example, in Figure 3 electrode pair 22 is shown to be arranged at a periphery of first group of piezoelectric nanofibers 18a. Electrode pair 24 is shown to be arranged at a periphery of the second group of piezoelectric nanofibers 18b. Electrode pair 26 is shown to be arranged at a periphery of third group of piezoelectric nanofibers 18c. Electrode pair 28 is shown to be arranged at a periphery of fourth group of piezoelectric nanofibers 18d. Each piezoelectric nanofiber group may therefore have a corresponding outer electrode pair from which the output voltage can be measured.

[0083] The number of outer electrode pairs may equal the number of groups of piezoelectric nanofibers. Although only four groups of piezoelectric nanofiber groups I outer electrode pairs are shown, it will be appreciated that any number of groups / electrode pairs can be used.

[0084] In examples, a number of pairs of the outer electrodes 16 is equal to a number of the piezoelectric nanofibers 18, and each piezoelectric nanofiber of the plurality of piezoelectric nanofibers 18 extends to a respective outer electrode pair 16. As such, there may be a one-to-one correspondence between each piezoelectric nanofiber and outer electrode pair, providing for accurate sound content and / or direction discrimination based on the generated electrical signal(s). In other examples, a few of the piezoelectric nanofibers of the plurality of piezoelectric nanofibers 18 extend to a respective outer electrode pair 16, such as two, three, four, five or six nanofibers. In general, any number of piezoelectric nanofibers may extend to a respective outer electrode pair 16.

[0085] In the examples illustrated by Figures 1 to 3, the length of the plurality of piezoelectric nanofibers 18 decreases in a circumferential direction of the free end 12 of the cantilever 10, such that the plurality of piezoelectric nanofibers 18 form a spiral-like shape. The spiral-like shape of the piezoelectric nanofibers 18 results in the generation of an electrical signal from which sound direction and / or content can be determined with relatively high accuracy.

[0086] In particular, the length of the plurality of piezoelectric nanofibers 18 may decreases stepwise in the circumferential direction of the free end 12 of the cantilever 10. The stepwise design allows for simplified and more controlled signal transmission, enabling clearer analysis of individual channel effects. This modular approach facilitates detailed study and optimization of each discrete output channel, which is particularly valuable in early-stage development and testing. Moreover, the stepwise structure can be scaled to include more channels or evolve toward a quasi-continuous configuration. Preferably, the size of each individual channel in this design is on the nanometre scale, offering high spatial resolution and precision.

[0087] Additionally or alternatively, the length of the plurality of piezoelectric nanofibers 18 may decrease continuously in the circumferential direction of the free end 12 of the cantilever 10. Therefore, different spiral-like arrangements of the piezoelectric nanofibers 18 can be provided.

[0088] The plurality of piezoelectric nanofibers 18 may comprise piezoelectric polymer nanofibers comprising piezoelectric nanoparticles. For example, the piezoelectric polymer may comprise one or more of poly(vinylidene fluoride-co-trifluoroethylene) copolymer (PVDF-TrFE) and polyvinylidene fluoride (PVDF). The piezoelectric nanoparticles may comprise barium titanate (BTO) nanoparticles. In such an example, the mechanical properties of the nanofibers 18 can be influenced by the dosage of BTO nanoparticles. For example, the BTO nanoparticles may be present at a concentration of 0-12 wt% relative to the weight of the PVDF-TrFE or PVDF, such as 0.1-12 wt% relative to the weight of the PVDF-TrFE or PVDF. This provides an optimum balance between the apparent modulus, strength and stain of the nanofibers 18.Preferably, the BTO nanoparticles are provided at a concentration of approximately 6 wt% relative to the weight of the PVDF-TrFE or PVDF, which provides the most optimum mechanical properties for the nanofibers.

[0089] Any of the example acoustic devices discussed herein may be self-powered acoustic devices, which do not require an external energy source to operate.

[0090] Referring now to Figure 4, there is shown an example system 100 according to an example of the present disclosure. The system 100 is shown to comprise the acoustic device 2 depicted in Figure 1. However, the system 100 may comprise any of the example acoustic devices discussed above.

[0091] The system 100 is configured to communicate with one or more data processors 30. In the present example, the system 100 comprises the one or more data processors 30.

[0092] In examples, the one or more data processors may be integrated within the acoustic device 2 for on-device processing of the electrical signal generated by the plurality of piezoelectric nanofibers 18. In other examples, the one or more data processors 30 may be implemented on one or more computers external to the acoustic device 2. In further examples, the one or more data processors 30 may be implemented in a cloud-based processing system, and a wireless unit in the acoustic device 2 may transmit the electrical signal to the one or more data processors 30.

[0093] In the present example, the one or more data processors 30 are configured to receive the electrical signal, generated by the plurality of piezoelectric nanofibers 18, from the acoustic device 2. For example, the electrical signal may be transmitted to the one or more data processors 30 via electrical cabling. In other examples, the electrical signal may be transmitted to the one or more data processors 30 wirelessly via a wireless unit.

[0094] The one or more data processors 30 are configured to process the electrical signal using a machine learning model to determine at least one of a direction and content of the acoustic signal based on features of the electrical signal. As discussed above, the features of the electrical signal correspond to spectral features of the measured voltage output across the inner pair of electrodes 14 and / or the outer pair of electrodes 16. The machine learning model is configured to learn and identify the specific spectral features of voltage signals collected as the location of the sound source is changed. The model is used to capture the essential nonlinearities of the acoustic features transduced by the vibration of the asymmetricpiezoelectric nanofibers 18 in the acoustic device 2, which provides for high accuracy in sound direction recognition and sound content decodification. The machine learning model may comprise an artificial neural network, for example.

[0095] As shown in Figure 5, the machine learning model is configured to analyse essential directional features from the electrical signals generated by the acoustic device 2. Specifically, Figure 5 illustrates a schematic workflow of using the system 100 for recognising sound direction. It will be appreciated that an analogous workflow can be applied to recognise sound content.

[0096] In particular, the data acquisition phase illustrated in Figure 5 involves the rotation of a mouth simulator around the acoustic device, generating sound sweeps at various angles to the device in both horizontal and vertical planes. The pre-processing phase shown in Figure 5 includes detrending, splitting and / or applying a short-time Fourier transform (STFT) to the electrical signal transmitted by the acoustic device. The STFT spectra may then be merged as input for training and testing to the machine learning model. The machine learning model may then be used in the final phase shown in Figure 5 to predict a direction and / or content of the sound. The result may then be stored in a known direction / content database in the machine learning model training phase, or output to a user device in order to communicate the results of the directions or sound contents to a user in the testing phase.

[0097] Figure 6 schematically illustrates a structure of a machine learning model according to an example of the present disclosure. As shown in Figure 6, the machine learning model may be structured using machine learning algorithms comprising an auditory spectrum encoder 32 and a sound feature extractor 34. The auditory spectrum encoder 32 is configured to encode the input combined spectra 33 as shallow features. The auditory spectrum encoder 32 may comprise one or more convolutional neural networks. In the example shown in Figure 6, the auditory spectrum encoder 32 comprises a plurality of convolutional layers 35a-d. A residual connection 36 comprising the plurality of convolutional layers is set in each residual network block 38a-n to alleviate auditory information loss during feature encoding.

[0098] The sound feature extractor 34 is configured to exploit attention mechanism to deal with deep auditory features, and comprises a formal attention layer 40 and an auditory transformer layer 42. This helps the model to focus more on the important STFT channelled information by setting different feature weights. The auditory transformer layer 42 comprises normalisation layers 44, 48 configured to normalise the input across the feature dimension. This stabilises and accelerates training by reducing internal covariate shift. The auditory transformer layer 42further comprises multi-head attention 46. The multi-head attention 46 allows the model to attend to information from different positions in the input sequence simultaneously, capturing temporal and contextual relationships in auditory signals. The auditory transformer layer 42 further comprise a multi-layer perceptron 50 configured to process the attention output. The multi-layer perceptron 50 enables complex feature transformations and enhances the model's representation capacity. Each component of the auditory transformer layer 42 is configured to work together to extract, refine, and encode meaningful patterns from auditory input for downstream tasks. For example, an output from the sound feature extractor 34 can be used to provide a prediction 52 of sound direction and / or content.

[0099] For applying classification-based prediction, the final layer of the machine learning model is a linear layer with n nodes which can provide the confidence of each category, where n refers to the number of data categories. Then the category with largest confidence is the final prediction result. The training loss of this strategy may be cross entropy loss. Loss of classification-based prediction can be defined as:

[0100] 1n

[0101] =n Z_i yt *los(Ti)

[0102]

[0103] i=l

[0104] where Lcisrefers to the loss of classification-based prediction, n is the number of data ytis the true label of data and ytis the predicted label.

[0105] For applying regression-based prediction, the final layer of the machine learning model is a linear layer which can directly provide the predictive value. The training loss of this strategy may be mean squared error loss. Loss of regression-based prediction can be defined as:

[0106] 1

[0107] ^rea =~yd2

[0108]

[0109] i=l

[0110] where Lregrefers to the loss of regression-based prediction, k is the number of training samples, ytis the true value of data and ytis the predicted value.

[0111] Therefore, to process the electrical signal generated by the acoustic device 2, the one or more data processors 30 may be configured to extract features associated with at least one of the direction and content of the acoustic signal from the electrical signal; input the extracted features into the machine learning model trained for at least one of sound direction and content classification and regression; and determine at least one of the direction and content of theacoustic signal based on an output from the machine learning model. Such an approach is consistent with that shown in Figure 5.

[0112] Furthermore, as shown in Figure 5 and as described above, the one or more data processors 30 may be configured to pre-process the electrical signal prior to processing the electrical signal using the machine learning model. To pre-process the electrical signal, the one or more data processors 30 may be configured to at least one of: detrend the electrical signal for input to the machine learning model; split the electrical signal for input to the machine learning model; and apply a STFT function to the electrical signal to generate a plurality of STFT signals and / or merge the STFT signals for input to the machine learning model.

[0113] As also described above, the machine learning model may be trained based on historical electrical signals generated by the acoustic device as shown in Figure 5. For example, sound may be directed to the acoustic device 2, which generates an electrical signal in response. The electrical signal (i.e. the voltage signal) may then be pre-processed and input as training data into machine learning model.

[0114] Figure 7 schematically represents a method of determining at least one of a direction and content of an acoustic signal, according to an embodiment. The method comprises a first step S1 of receiving, by an acoustic device comprising a plurality of asymmetrically arranged piezoelectric nanofibers, an acoustic signal. For example, the acoustic device may be any of the example acoustic devices discussed above, such as acoustic device 2.

[0115] The method further comprises a second step S2 of converting, by the acoustic device and using the plurality of asymmetrically arranged piezoelectric nanofibers, the acoustic signal to an electrical signal comprising features associated with at least one of a direction and content of the acoustic signal. The piezoelectric nanofibers convert the acoustic signal to the electrical signal by virtue of the piezoelectric effect, in the same manner discussed above for acoustic device 2.

[0116] The method further comprises a third step S3 of outputting, by the acoustic device and to one or more data processors, the electrical signal. The method also comprises a fourth S4 of receiving, by one or more data processors, the electrical signal. For example, the electrical signal may be output by the acoustic device to the one or more data processors via electrical cabling, or wireless transmitter / receiver units.The method comprises a fifth step S5 of processing, by the one or more data processors, the electrical signal using a machine learning model to determine at least one of the direction and content of the acoustic signal based on the features of the electrical signal. The processing may correspond to any of the processing discussed above and performed by the one or more data processors 30.

[0117] Optionally, the method may include a further step of outputting the content and / or direction of the acoustic signal to a user interface, to communicate the results of the sound content and / or direction determination to a user.

[0118] It will be appreciated that the method may comprise one or more additional steps corresponding to the functionality performed by the acoustic device 2 and / or the one or more data processors 30, such as any of the actions taken by the acoustic device 2 and or the one or more data processors 30 discussed previously.

[0119] Piezo-acoustic performance of the acoustic device

[0120] Figures 8a-e show plots of experimental data obtained using an acoustic device similar to that shown in Figure 1. The acoustic device used in this example has four output channels (corresponding to each group of piezoelectric nanofibers) in varying lengths, as shown in Figure 8b. Specifically, in this example the radial distance between the free end of the cantilever and the inner boundary of the body is: 15 mm in Channel 1; 20 mm in Channel 2; 25 mm in Channel 3; and 30 mm in Channel 4. Figures 8a and 8c present the vibration displacements and voltage outputs of each channel, respectively. The solid dots highlighted in Figure 8b represent the monitored vibration displacement of each channel's representative center point. The resonant behavior of each channel in the acoustic device is distinct. Channel 1 exhibits a resonant frequency of 655 Hz, while Channel 2, 3, and 4 resonate at frequencies of 310 Hz, 210 Hz, and 190 Hz, respectively.

[0121] Piezo-acoustic signal analysis suggests that the voltage output of each channel is influenced by its neighbouring channels. For example, the resonant frequency of Channel 4 at 190 Hz overlaps with a shoulder peak in Channel 3, whose resonant frequency is at 342 Hz and is also reflected in Channel 2. Additionally, a resonance peak at 120 Hz is observed in both displacement and voltage outputs across all channels, which can be attributed to the resonant vibration of the inner cantilever. The short-time Fourier transform (STFT) of the displacement at the representative points, stimulated by a sound frequency sweep, reveals the timefrequency distributions (as shown in Figure 8b). Vibrational peaks are observed at low frequencies in Channels 4 and 3, which have relatively larger radii and longer spring lengths.The resonant peaks shift to higher frequencies as the spring lengths decrease in Channels 2 and 1.

[0122] The acoustic device in this example features four separate channels that can also be connected in series to provide a single channel output, where the upper or bottom electrodes are physically connected accordingly. To construct and analyze the tonotopic profile of the device in the present example, the piezo-acoustic output of the single-channel configuration was recorded under a frequency sweep scanning. The frequency characteristics of the acoustic device were then analysed using STFT and peak analysis, as depicted in Figure 8d. In the frequency analysis, voltage peaks were identified at frequencies of 655 Hz, 374 Hz, 240 Hz, 188 Hz, 142 Hz, and 93 Hz. These peak frequencies were further analysed by fitting Gaussian functions. The piezo-acoustic peak of the single channel output of the acoustic device aligns with the resonant peaks observed in each separate channel. This indicates that the acoustic device is capable of transducing and summarising detailed sound information by capturing features across a wider range of frequencies through all four channels. This remarkable capability is attributed to the utilization of highly sensitive asymmetrically arranged piezoelectric nanofibers and the establishment of multiresonant coupling conditions.

[0123] To gain deeper insights into the coupled multiresonant behaviours of the acoustic device, the out-of-plane vibration was characterised using a laser vibrometer. The scanning point density was set to approximately 1.1 mm. The experimental mapping of the acoustic device, as shown in Figure 8e, demonstrates clear separation of vibrational peaks along the radial directions of the nanofibers. In higher resonance modes, the vibrational modes observed in the experimental mapping became more intricate, as depicted in Figure 8e. Specifically, the maximum displacement was observed at a low frequency of around 180 Hz in Channel 4. As the frequency increased to 200 Hz, the maximum displacement moved to Channel 3. As the frequency further increased to 320 Hz, 420 Hz, and 660 Hz, additional vibrational peaks emerged along the radial direction (highlighted by dash loops), echoing the displacements measured at the centre points of each channel (as shown in Figures 8a and 8b).

[0124] Clearly, the spiral shape and asymmetric piezo-nanofibers uniquely define the spectra of vibration and corresponding voltage output of the acoustic device in response to varying frequency, reminiscent of tonotopic function of the human auditory system.

[0125] Machine based sound direction and contentFigures 9a-d show further plots of experimental data obtained using an acoustic device according to an example of the present disclosure.

[0126] The piezoelectric acoustic devices discussed herein have been demonstrated to exhibit outstanding multiresonant sound transduction, showcasing superior sensitivity and mimicking the tonotopic profile of the human auditory system. The piezo-acoustic signal transduction carries rich information correlated to sound content and direction, which effectively emulate the functional information of both “peripherical and central auditory system”. With the assistance of machine learning, an artificial "auditory cortex" model can be developed into the system for a fully functioning hearing system.

[0127] An attention-based machine learning model was designed to analyse essential directional features from electrical signals generated by the acoustic device, as discussed above in respect of Figure 5. For each direction, at least 200 piezo-acoustic repeated scans were collected and used as inputs fortraining and testing procedures. After 150 iterations, the model achieved an accuracy of over 90% on both training and validation data, indicating successful learning of patterns and effective generalization to unseen instances

[0128] During the “hearing training” procedure shown in Figure 5, all training data was fed into the model by 200 iterations, with an additional 20 untrained data used for validation and 20 data points used for testing. The trained model was combined with the acoustic device as a whole artificial auditory module (e.g. system 100 illustrated in Figure 4). The “hearing testing” procedure shown in Figure 5 aims to evaluate the performance of the trained model, where the recognition was performed by comparing the unknown sound input features with the pretrained knowledge. The recognition accuracy is the ratio of correct predictions to the total testing number. The module exhibited exceptional accuracy in direction recognition along the polar axis in three dimensions, covering a solid angle of 4TT.

[0129] Figures 9a and 9b present the classification-based recognition accuracy on the xy- and yz-planes, respectively, with an azimuthal step of 30°. The average accuracy for all 12 angles in the xy-plane (0) was 97±3.2%, while in the y plane (<t>), it was 92.1 ±8.1%. Moreover, the recognition accuracy for different distances along the z-axis reached 100%, indicating the device's sensitivity to changes in distance.

[0130] To investigate the effect of channel number on the recognition accuracy, the training data was collected from an acoustic device with 1 channel, 2 channels and 4 channels, which physically combined the electrodes as illustrated Figure 9c. The azimuthal step precision of the modulewas investigated by training and testing with different stepwise angles of 1°, 5°, 10°, 15° and 30°. The results show that the optimal accuracy of stepwise angles reduces when the moving step is smaller than 10°. Notably, the accuracy reduces more significantly for the configurations with less channels. The recognition accuracy is always higher when using the acoustic device with higher number of channels. In the 2 channels configurations, it can be also noticed that the crossed channel configuration (X) shows a higher recognition accuracy than that of paralleled channels ( / / ). This is related to the fact that the vibration of one channel affects the neighbouring channels. The crossed channels reduced this effect by using the opposite channels, which was also one of the reasons to choose the attention based model. A demo of the real-time recognition of sound directions was demonstrated with a pre-trained acoustic device module. The rotational stage travelled to a random angle and played the sound to stimulate the acoustic device. A recognition was made and displayed on a user interface screen. The overall accuracy of the real-time sound direction recognition using this approach was found to be very high at 96.0% .

[0131] The above-described approach excels in classifying labelled sound directions. However, it may be difficult to cover the complete range of sound direction. Thus, a regression-based prediction is also provided to cover the complete range of the sound direction. The labels indicating the position of the sound source (0, 30, 60...330) were hidden before feeding the data to the model. The box plot in Figure 9d shows the performance of regression-based predictions, where 477 unlabelled data from the xy-plane were used to test the model and request for a prediction of any angles between 0°-360°. Notably, the labels of these sound directions are distributed close to their exact direction angles. The R-squared value of the results is 0.93, which also proves the good training quality of regression-based prediction. To prove the self-learning ability of the model, it is expected to recognise the angles that have never been labelled or trained, such as 22°, 40° and 45°. The piezo-acoustic signals were collected from these unknown, untrained, and unlabelled directions and required the prediction of their detailed directions. The regression-based model responded with reasonable predictions that were close to the exact angles (as shown in Figure 9d inset).

[0132] A system has therefore been described which combines a programmable asymmetric structure of piezoelectric nanofibers with digitally intelligent neural networks, resembling the processing stages of the human auditory system. An acoustic device described herein, based on asymmetrically arranged piezoelectric nanofibers, demonstrates exceptional functions in sound recognition and directional recognition with excellent sensitivity and accuracy. The asymmetric spiral geometry of the piezoelectric nanofibers not only exhibits complex multi-resonant vibration spectra, effectively mimicking the vibrational transduction and frequencyseparation of the basilar membrane in the human cochlea, but also mimics and even outperforms the spatial intelligence of human brain in assisting with neural networks. The system demonstrates outstanding accuracy in recognizing sound in all directions.

[0133] The example acoustic devices discussed herein can be applied as a wearable or even implantable hearing device for hearing loss patients, offering enhanced auditory experiences for those with hearing impairments. Other applications include use of the acoustic device in smart robotics, or voice interfaces like artificial cochlea for human-machine interfaces. Moreover, the machine-learning model can provide a tool for studying the nonlinear signal processing of complex sound from the inner ear to the brain.

Claims

Claims1. An acoustic device for sound direction and content recognition, comprising:a body comprising a gap defined by an inner boundary of the body;a cantilever extending into the gap from the inner boundary and comprising a free end in the gap;at least one pair of inner electrodes arranged at the free end of the cantilever; at least one pair of outer electrodes arranged at the inner boundary of the body; and a plurality of piezoelectric nanofibers extending radially from the free end of the cantilever to the inner boundary and arranged asymmetrically around the free end,wherein the piezoelectric nanofibers are configured to convert an acoustic signal into an electrical signal detectable using the at least one pair of inner electrodes and / or the at least one pair of outer electrodes.

2. The acoustic device according to claim 1, wherein the electrical signal comprises features associated with a direction and content of the acoustic signal.

3. The acoustic device according to claim 1 or 2, wherein:the plurality of piezoelectric nanofibers comprises a plurality of groups of piezoelectric nanofibers;the length of each piezoelectric nanofiber in a respective group is the same; and the length of the piezoelectric nanofibers in each group is different from the length of the piezoelectric nanofibers in each other group.

4. The acoustic device according to claim 3, wherein each group of piezoelectric nanofibers is tuned to a resonant frequency of the respective group of piezoelectric nanofibers.

5. The acoustic device according to claim 4, wherein each group of piezoelectric nanofibers is tuned to a different resonant frequency.

6. The acoustic device according to any of claims 3 to 5, wherein:the acoustic device comprises a plurality of outer electrode pairs arranged at the inner boundary of the body; andeach outer electrode pair of the plurality of outer electrode pairs is arranged at a periphery of a respective group of piezoelectric nanofibers.

7. The acoustic device according to any preceding claim, wherein the length of the plurality of piezoelectric nanofibers decreases in a circumferential direction of the free end of the cantilever such that the plurality of piezoelectric nanofibers form a spiral-like shape.

8. The acoustic device according to claim 7, wherein the length of the plurality of piezoelectric nanofibers decreases stepwise in the circumferential direction.

9. The acoustic device according to claim 7, wherein the length of the plurality of piezoelectric nanofibers decreases continuously in the circumferential direction.

10. The acoustic device according to any preceding claim, wherein:a number of pairs of the outer electrodes is equal to a number of the piezoelectric nanofibers; andeach piezoelectric nanofiber of the plurality of piezoelectric nanofibers extends to a respective outer electrode pair.

11. The acoustic device according to any preceding claim, wherein the body further comprises a polymer layer between each electrode of the at least one pair of inner electrodes, and between each electrode of the at least one pair of outer electrodes.

12. The acoustic device according to any preceding claim, wherein the acoustic device is a self-powered acoustic device.

13. The acoustic device according to any preceding claim, wherein the plurality of piezoelectric nanofibers are piezoelectric polymer nanofibers comprising piezoelectric nanoparticles.

14. The acoustic device according to claim 13, wherein:the piezoelectric polymer is poly(vinylidene fluoride-co-trifluoroethylene) copolymer, PVDF-TrFE, or polyvinylidene fluoride, PVDF;the piezoelectric nanoparticles comprise barium titanate, BTO, nanoparticles; and a concentration of the BTO nanoparticles is 0-12 wt% relative to the weight of the PVDF-TrFE or PVDF.

15. A system comprising the acoustic device according to any preceding claim, wherein the system is configured to communicate with one or more data processors configured to:receive the electrical signal from the acoustic device; andprocess the electrical signal using a machine learning model to determine at least one of a direction and content of the acoustic signal based on features of the electrical signal.

16. The system according to claim 15, wherein to process the electrical signal, the one or more data processors are configured to:extract features associated with at least one of the direction and content of the acoustic signal from the electrical signal;input the extracted features into the machine learning model trained for at least one of sound direction and content classification and regression; anddetermine at least one of the direction and content of the acoustic signal based on an output from the machine learning model.

17. The system according to claim 15 or 16, wherein:the one or more data processors are configured to pre-process the electrical signal prior to processing the electrical signal using the machine learning model; andto pre-process the electrical signal, the one or more data processors are configured to at least one of:detrend the electrical signal for input to the machine learning model; split the electrical signal for input to the machine learning model; and apply a short-time Fourier transform, STFT, function to the electrical signal to generate a plurality of STFT signals and / or merge the STFT signals for input to the machine learning model.

18. The system according to any of claims 15 to 17, wherein the machine learning model is trained based on historical electrical signals generated by the acoustic device.

19. A method of determining at least one of a direction and content of an acoustic signal, the method comprising:receiving, by the acoustic device according to any of claims 1 to 14, an acoustic signal; converting, by the acoustic device and using the plurality of asymmetrically arranged piezoelectric nanofibers, the acoustic signal to an electrical signal comprising features associated with at least one of a direction and content of the acoustic signal;outputting, by the acoustic device and to one or more data processors, the electrical signal;receiving, by one or more data processors, the electrical signal; andprocessing, by the one or more data processors, the electrical signal using a machine learning model to determine at least one of the direction and content of the acoustic signal based on the features of the electrical signal.

20. The method according to claim 19, wherein the processing comprises:extracting features associated with at least one of the direction and content of the acoustic signal from the electrical signal;inputting the extracted features into the machine learning model trained for at least one of sound direction and content classification and regression; anddetermining at least one of the direction and content of the acoustic signal based on an output from the machine learning model.

21. The method according to any of claims 19 to 20, further comprising:pre-processing, by the one or more data processors, the electrical signal prior to processing the electrical signal using the machine learning model, wherein the pre-processing comprises at least one of:detrending the electrical signal for input to the machine learning model; splitting the electrical signal for input to the machine learning model; and applying a Short-time Fourier transform, STFT, function to the electrical signal to generate a plurality of STFT signals and / or merge the STFT signals for input to the machine learning model.

22. The method according to any of claims 19 to 21 , wherein the machine learning model is trained based on historical electrical signals generated by the acoustic device.

23. A method of manufacturing the acoustic device according to any of claims 1 to 14, the method comprising:attaching the at least one pair of inner electrodes onto a lower layer and an upper layer of the body, the upper layer and lower layer comprising the gap;attaching the at least one pair of outer electrodes to the lower layer and the upper layer;electrospinning the plurality of piezoelectric nanofibers into the gap of the lower layer; andjoining the upper layer and lower layer together.

24. The method according to claim 23, further comprising:applying the upper layer of the body and the lower layer of the body on opposing sides of a polymer layer.

Images