Gradient-distributed metal-halide dynamic memristors for adaptive and robust voiceprint recognition

Abstract

Inspired by the auditory system’s capacity to process spatiotemporal sound patterns, voiceprint recognition plays a vital role in identity authentication and security. However, current platforms often face challenges of speech frequency and amplitude variability, hindering accurate feature extraction in noisy environments. To address these issues, a large-scale hybrid metal-halide dynamic memristor (MHDM) featuring an engineered gradient-distributed architecture is developed for adaptive voiceprint recognition. The spontaneously graded metal-halide functional layer allows for precise modulation of Schottky barriers and redistribution of interface charges. This design achieves µs-scale response, enhances noise tolerance (over 20% improvement in signal-to-noise ratio), and enables kHz-scale dynamic signal processing. Experimental results demonstrate that the MHDM achieves a voiceprint recognition accuracy of 99.3%, maintaining high performance at 93.2% even in realistic background noise. These findings demonstrate the system’s potential for secure and efficient voiceprint recognition, combining scalability with robust performance in noisy environments.

Voiceprint recognition is critical for biometric security, yet it remains susceptible to real-world acoustic disruption. Shao et al. report a gradient-distributed metal-halide memristor that exploits intrinsic adaptive dynamics to achieve robust and accurate voice recognition under noisy conditions.

Introduction

Voiceprint recognition seeks to emulate the biological auditory system’s ability to extract robust identity signatures from dynamic, contaminated acoustic streams^1–3. The development of efficient hardware systems for this task is limited by the structural separation of signal acquisition and processing units in conventional architectures^4,5. This conventional design lacks adaptive signal preprocessing at the sensory interface and fails to provide the inherent noise suppression required for reliable operation in real-world conditions^6,7. Although digital implementations can execute advanced denoising algorithms, their serial processing nature and high power consumption are incompatible with the strict latency and energy constraints of edge computing applications^8–10. Furthermore, existing sensing hardware remains largely passive and lacks the local processing capabilities necessary for on-site feature enhancement.

While neuromorphic memristive devices offer a potential hardware platform for integrated spatiotemporal processing, their practical implementation remains hindered by fundamental differences in switching properties^11,12. Conventional threshold switches and filamentary memristors exhibit abrupt, inherently binary switching behavior^13,14. These mechanisms pose some challenges to the continuous analog state modulation required for accurate encoding of time-varying vocal features^15,16. Even among volatile memristors, imprecise dynamic transients often cause information loss during the transduction of complex auditory patterns, particularly affecting subtle acoustic features¹⁷. Furthermore, conventional ion-migration-based devices suffer from progressive interfacial charge accumulation during sustained operation^10,18. This parasitic memory effect is particularly detrimental in auditory tasks where high-frequency acoustic stimuli drive the device into uncontrolled conductance variation. Such a transition disrupts the usable dynamic range and blinds the system to the subtle spectral variations that define a voiceprint, posing a barrier to real-world deployment.

Here, we report a large-scale hybrid metal-halide dynamic memristor (MHDM) featuring a gradient-distributed functional layer designed for noise-robust voiceprint recognition. The compositional gradient along the film thickness arises spontaneously during one-step thermal evaporation of metal halide precursors. By synergistically coupling ion migration dynamics with charge transport mechanisms, the device achieves adaptive modulation of Schottky barrier characteristics and interfacial charge distribution. This hybrid architecture leverages dimensional confinement to suppress uncontrolled ion accumulation^19–22, ensuring exceptional operational stability and processing adaptability for complex, time-varying speech signals. With a fast signal attenuation time of less than 10 μs, the device is specifically optimized for high-frequency temporal processing. In practical evaluations using a diverse multi-speaker corpus, the system achieves 99.3% accuracy. Under real-world noisy conditions, the system maintains a robust recognition accuracy of 93.2% while simultaneously improving the signal-to-noise ratio (SNR) by over 20%, positioning it as a promising candidate for hardware-based forensic and security applications.

Results

Figure 1 illustrates the architecture and operational workflow of the voiceprint recognition system for identity authentication. As shown in Fig. 1a, the system enables remote non-contact voice authentication via mobile devices under challenging environmental conditions, even when background noise masks key vocal features. Figure 1b presents the device structure, which consists of a vertically stacked architecture with a compositionally graded active layer formed via one-step thermal evaporation. Upon continuous voltage stimulation, this gradient architecture enables adaptive signal processing through voltage-driven ionic migration, dynamically modulating the interfacial charge distribution in response to sequential input stimuli (Fig. 1c). The resulting adaptive current response facilitates real-time signal acquisition and dynamic feedback, thereby capturing subtle inter-speaker voiceprint variations. The acquired signals undergo intrinsic temporal filtering and adaptive dynamics within the MHDM, effectively attenuating environmental interference while preserving critical vocal features. The processed signals are subsequently routed to a neural network for speaker classification and recognition (Fig. 1d). This design provides noise immunity, improved accuracy, and reliability through real-time feature adaptation in complex auditory environments.

Fig. 1. Architecture of the voiceprint recognition system for identity authentication.

a Schematic of remote non-contact voice authentication via mobile phone under various environmental noise conditions. Background noises mix with voice signals and obscure key vocal features. b Structure of the MHDM based on graded Cs-Cu-I. c Adaptive current response characteristics of the MHDM under electrical stimulation. d Processed signals are delivered to a neural network for speaker classification and recognition.

Composition-gradient engineered memristor

The MHDM adopts a vertically stacked ITO/Cs-Cu-I/Au architecture (Fig. 2a). Its active Cs-Cu-I layer, approximately 80 nm thick as confirmed by cross-sectional scanning electron microscope (SEM) (Supplementary Fig. 1), was deposited via a precisely controlled thermal co-evaporation process. This vacuum-based method was selected owing to its controllability in forming uniform, pinhole-free films, a result confirmed by atomic force microscopy (AFM) (Supplementary Fig. 2). This approach offers advantages over conventional solution processing, which is often compromised by the interference of anti-solvent and spin-coating processes^23,24. Fig. 2b shows a photograph of the fabricated large-scale crossbar structure, consisting of 64 × 64 devices on a 10 cm × 10 cm (≈ 4 inch) substrate, demonstrating the high integrability and scalability of the presented platform. Statistical analysis of 100 devices across the array (Supplementary Fig. 3) confirms consistent device-to-device uniformity, as evidenced by the limited variability in key electrical parameters. This result indicates the robustness and reproducibility of the vacuum-based fabrication process, suggesting its suitability for large-scale hardware integration.

Fig. 2. Electrical characteristics of the MHDM.

a Schematic illustration of the MHDM structure. b Photograph of a 10 cm × 10 cm crossbar array containing 64 × 64 devices. c Typical I–V characteristics of the MHDM under consecutive bipolar voltage sweeps, showing pinched hysteresis loops indicative of analog memristive behavior. d, e STP responses induced by single electrical pulses with pulse widths of 50 μs (d) and 100 ms (e) at a fixed pulse amplitude of 0.3 V (V_R = 0.05 V). Insets show the corresponding current decay upon pulse removal. f STP responses under 50 consecutive pulses at frequencies of 9.99 Hz, 99 Hz, and 5000 Hz (P_A = 0.3 V, V_R = 0.05 V). Right panels compare the evolution of read current under 9.99 Hz (potentiation) and 5000 Hz (depression) stimulation. g Temperature-dependent STP behavior under 50 consecutive 0.3 V pulses and 0.5 V pulses at 99 Hz, measured at room temperature (RT) and elevated temperature (373.15 K).

The MHDM exhibits analog memristive behavior, as shown in Fig. 2c and Supplementary Fig. 4, which presents the typical current-voltage (I–V) characteristics under repeated bipolar voltage sweeps. The device displays clear pinched hysteresis loops under bipolar periodic electrical stimulation, a feature characteristic of analog memristors^25,26. The asymmetric I–V characteristics suggest the presence of a Schottky barrier at the Cs-Cu-I/electrode interfaces. To investigate the temporal dynamics of the device, transient electrical measurements were conducted by applying a single electrical pulse. As shown in Fig. 2d, e, the current responses were recorded under pulses with increasing pulse widths (P_W) of 50 μs and 100 ms, while maintaining a constant pulse amplitude (P_A) of 0.3 V. The current increased rapidly with pulse application and returned to baseline immediately upon its removal. This reversible behavior closely resembles the short-term potentiation (STP) observed in biological synapses^27,28. Analysis of the current decay under a single 0.3 V pulse reveals a biexponential relaxation with a dominant time constant of 7.16 μs for a 50 μs pulse width, which marginally increases to 7.28 μs when the width is extended to 100 ms. This biexponential relaxation arises from two concurrent physical processes. Specifically, the fast component originates from the rapid detrapping and discharge of carriers at or near the metal-halide interface. In contrast, the slower component is mediated by the back-diffusion of ions away from the Cs-Cu-I/Au interface. Supplementary Fig. 5 provides additional insights into the dynamic interplay between charge accumulation and relaxation under varied pulse conditions. Each positive voltage pulse induces an abrupt current spike followed by rapid relaxation to the baseline, consistent with STP behavior. This dynamic response is maintained across various operational parameters, where higher P_A and P_W values produce larger peak currents while preserving the characteristic decay behavior.

Frequency and amplitude effects on adaptive MHDM dynamics

The frequency-dependent response characteristics were systematically investigated by applying 50 consecutive pulses (P_A = 0.3 V) across different frequencies (9.99, 99, and 5000 Hz), with the voltage sequence detailed in Supplementary Fig. 6. Figure 2f illustrates distinct frequency-dependent behavior, where 9.99 Hz stimulation induces pronounced peak current potentiation, which gradually diminishes at 5000 Hz. The successive pulses at 9.99 Hz induce a cumulative enhancement in read current, demonstrating potentiation, whereas stimulation at 5000 Hz leads to progressive suppression, indicative of depression behavior. Supplementary Fig. 7 further compares device responses under 50 consecutive 0.3 V pulses with P_W of 1 ms and 0.1 ms, highlighting the significant impact of frequency and P_W on dynamic performance. Notably, high-frequency stimulation can also lead to a gradual decrease in peak current, as shown in Supplementary Fig. 8, where peak current declines with increasing frequency across all tested pulse widths. A clear reduction is observed at 5000 Hz and 6666 Hz, emphasizing the frequency-dependent transition from potentiation to depression.

The effect of pulse amplitude on device performance was systematically explored using 10 ms positive pulses at 0.3 and 0.5 V. The voltage sequence is detailed in Supplementary Fig. 9. At 0.3 V, the device exhibits a gradual increase in peak current, reflecting potentiation driven by cumulative conductance modulation. In contrast, at 0.5 V, the peak current progressively declined, indicating a transition to depression (Fig. 2g). Temperature-dependent measurements provide further insight into the underlying mechanism. As shown in Fig. 2g, under identical pulsing conditions, both the potentiation and depression effects become more pronounced at elevated temperature (373.15 K). The enhanced potentiation at higher temperatures indicates more efficient ion redistribution under stimulation. At the same time, the more pronounced depression under high-frequency or high-amplitude pulses suggests accelerated ion accumulation at the interface. This thermal activation of both potentiation and depression indicates that the observed adaptive dynamics are governed by ion migration, which is intrinsically temperature-sensitive.

Engineering compositional gradients in Cs-Cu-I deposition

The frequency-dependent electronic performance of the device can be attributed to its distinct structural features in the active layer. As shown in Fig. 3a, the Cs-Cu-I film was fabricated via thermal co-evaporation using a mixed CsI and CuI source. The crystalline phase of the Cs-Cu-I film was further confirmed through X-ray diffraction (XRD) analysis, as shown in Supplementary Fig. 10. The diffraction peak at 26.20° corresponds to the (221) crystal planes of the orthorhombic CsCu₂I₃ phase^29,30. Additionally, a pronounced peak at 25.27° was attributed to the (111) planes of γ-CuI, indicating the presence of residual CuI within the film³¹. This excess accumulation of CuI can be attributed to its higher molar mass than CsI. During the initial deposition stage, co-evaporation of CsI and CuI leads to the formation of CsCu₂I₃, along with excess CsI, resulting in a Cs-rich region. Once CsI has mostly evaporated, CuI continues to deposit, forming a Cu-rich region. This sequential deposition process induces a vertical ionic gradient within the metal halide layer, forming interfacial barriers (Supplementary Note 1). This gradient distribution is further confirmed by cross-section elemental mapping of the device (Fig. 3b), which shows a relatively uniform distribution of Cu, strong Cs enrichment near the ITO electrode, and a clear gradient in I content across the film.

Fig. 3. Structural characterization and resistive switching mechanism of the device.

a A schematic of the evaporation deposition process for the Cs-Cu-I thin film. b Elemental mapping of the device cross-section; scale bar: 20 nm. c, d ToF-SIMS depth profiles (c) and corresponding 3D reconstructions (d) of the Cs-Cu-I layer within the memristor. e Schematic diagrams illustrate the proposed resistive switching mechanism of the memristor. f A schematic of the energy band diagram illustrating the barrier modulation caused by the accumulation of migrating V_I⁺ at the interface.

To further investigate the distribution of CuI within the CsCu₂I₃ films, time-of-flight secondary ion mass spectrometry (ToF-SIMS) was employed, focusing on the characteristic signals of Cu, Cs, and I (Fig. 3c, d). The results demonstrate distinct elemental distribution patterns with increasing sputter time. The Cs signal intensity rises rapidly and localizes primarily on the ITO region, while the Cu signal remains stable and concentrates near the Au. Additionally, the I signal derived from both CsI and CuI exhibits a gradient distribution, reflecting the dispersion of CsI and CuI within the film. Thus, a composite structure composed of CsCu₂I₃ and CuI formed spontaneously.

Based on the distinct compositional gradient observed within the film, the device’s working mechanism was further investigated. This gradient directly governs the ionic migration dynamics within the Cs-Cu-I film. As shown in Fig. 3e, under electrical stimulation, I⁻ ions are attracted toward the positively charged electrode, generating a large number of I⁻ vacancies (V_I⁺). When a positive voltage is applied to the ITO electrode, V_I⁺ are driven by the electric field and migrate towards the top Au electrode^32,33. Unlike conventional three-dimensional (3D) perovskites, which suffer from isotropic ionic diffusion and limited lattice stability, dimensional engineering has proven effective in stabilizing low-dimensional perovskites for optoelectronic applications. For instance, mixing two-phase architectures in quasi-two-dimensional (2D) perovskite can narrow phase distribution, eliminate van der Waals gaps, and enhance both film orientation and operational stability³⁴. Thus, our hybrid metal-halide architecture utilizes an engineered compositional gradient to implement dimensional confinement^35,36. This capability originates from the unique one-dimensional (1D) crystal structure of CsCu₂I₃, which consists of edge-sharing [CuI₄]³⁻ tetrahedra chains³⁷. The 1D structural units provide intrinsic, confined pathways for regulated ion transport along the crystalline channels, effectively suppressing the uncontrolled, random ion accumulation typically observed in 3D lattices.

To further elucidate the resistive switching mechanism, area-dependent electrical measurements were performed. As shown in Supplementary Fig. 11, the current scales with the effective electrode area across all conductance states, confirming that the switching behavior is governed by an interfacial effect rather than by the formation of metallic conductive filaments^38–40. To gain deeper insight into the charge-transport mechanism, temperature-dependent I–V characteristics were systematically investigated (Supplementary Fig. 12). The conductance increases consistently with increasing temperature, following an Arrhenius-type activation behavior^41,42. As extracted from Supplementary Fig. 13, the relatively low thermal activation energy E_a suggests that conduction is dominated by interfacial barrier-limited processes rather than filamentary metallic pathways⁴³. This is further supported by the linear relationship between ln (I) and V^1/2 observed under different temperatures, consistent with the Schottky emission mechanism. Furthermore, the dynamic evolution of device conductance is governed by dual-regime modulation of interfacial potential. Supplementary Fig. 14 reveals that the Schottky barrier height increases with applied voltage. This trend deviates from the classical behavior of the Richardson-Schottky model⁴⁴, suggesting that the barrier modulation in our device is additionally influenced by voltage-driven ionic redistribution. Such ionic migration modulates the local interfacial potential, contributing to the non-linear and hysteretic characteristics essential for analog memristive behavior.

Based on the ionic migration model, the device’s response to electrical pulses can be described as follows. The positive pulses applied to the ITO electrode cause V_I⁺ to migrate from the Cs-Cu-I layer toward the Cs-Cu-I/Au interface. This migration results in a progressive current increase upon initial continuous electrical stimulation, which can be attributed to the presence of the surface Cu-rich region that lowers the contact barrier with Au during the initial stage (Fig. 3f). During this initial phase, interfacial defects and redundant negative space charge are effectively neutralized by the accumulating V_I⁺, further reducing the contact barrier and thereby facilitating charge transport. Conversely, under sustained high-frequency and high-amplitude stimulation, intensified redistribution leads to excessive V_I⁺ accumulation at the interface, increasing the Schottky barrier height, which impedes hole transport. This adaptive barrier modulation manifests as an inhibitory response, enabling rapid depression. The gradient-engineered architecture thus synergistically couples ionic migration with interfacial charge transport, in which the field-driven redistribution of V_I⁺ at the top interface modulates carrier injection conditions. Such a mechanism enables the device to achieve an ultra-fast signal attenuation time while maintaining a stable recovery baseline, providing the physical foundation for the frequency-selective adaptation required for robust voiceprint recognition.

To further elucidate the role of CuI at the interface, it is crucial to evaluate its impact on charge transport and the interfacial barrier. A control device with a 40 nm functional layer (designed to avoid the formation of a Cu-rich layer) was fabricated without an initial ion gradient. Supplementary Fig. 15 illustrates the device structure and the corresponding XRD pattern of the CsCu₂I₃ film. When a 40 nm Cs-Cu-I film was deposited, only characteristic peaks of CsCu₂I₃ were observed, with no additional peaks from excess CuI, suggesting that CuI accumulation at the top layer primarily occurs during the latter deposition stages. The electrical characteristics of the device based on the proposed structure were subsequently systematically investigated. As illustrated in Supplementary Fig. 16 and Fig. 17, the device exhibits clear volatile behavior under single-electrical-pulse stimulation with varying P_A and P_W. Notably, the decay time increases proportionally with the P_A and P_W, indicating a voltage- and time-dependent relaxation mechanism. Furthermore, the current response under continuous pulse stimulation at different frequencies was investigated. As depicted in Supplementary Fig. 18, under consecutive 0.5 V pulses, the current shows pronounced potentiation without depression as the stimulation frequency increases, a behavior distinctly different from that of the MHDM. This observation further supports the conclusion that the accumulation of CuI creates an interfacial barrier that inhibits charge transport, thereby playing a crucial role in the observed current inhibition behavior^45,46. We further investigated the impact of the active layer thickness by fabricating devices with 120- and 160-nm-thick active layers. As shown in Supplementary Fig. 19, thicker films exhibit a reduced dynamic range and lower conductance levels. This performance degradation is attributed to weakened internal electric fields and extended ionic transport pathways, both of which hinder efficient Schottky barrier modulation at the interface. Therefore, the 80 nm thickness was identified as the optimal design parameter, providing a critical balance between maintaining a stable compositional gradient and sustaining the high-speed dynamics required for real-time acoustic processing.

To further decouple the individual contributions of the two phases and the gradient architecture, control devices consisting of a pure CuI layer (ITO/CuI/Au) were fabricated under identical deposition conditions. As shown in Supplementary Fig. 20, the device displays quasi-Ohmic conduction with high leakage current and negligible resistive switching or hysteresis. This indicates that neither material alone forms a Schottky contact strong enough for memristive switching. In contrast, the gradient ITO/Cs-Cu-I/Au architecture exhibits clear rectification and pinched hysteresis loops. This observation confirms that the gradient Cs-Cu-I is essential for establishing a modifiable Schottky barrier. In this configuration, CsCu₂I₃ serves as a reservoir of mobile V_I⁺, while the CuI layer provides an interface whose barrier height can be dynamically tuned by ion accumulation and depletion. The synergistic coupling between the two phases enables the adaptive memristive behavior observed in our devices.

Subsequent calculations of the ion migration energy, which facilitates V_I⁺ formation in both CsCu₂I₃ and CuI (Supplementary Fig. 21 and Fig. 22), reveal that the energy barrier for migration is significantly higher in CuI than in CsCu₂I₃. This indicates that, under low-amplitude voltage or low-frequency stimulation, migration predominantly occurs within CsCu₂I₃, where lower migration energy facilitates easier ion displacement. Furthermore, the deposition method used to form CsCu₂I₃ introduces additional ionic vacancies due to compositional differences among the evaporated components^47,48. In contrast, CuI exhibits a more complete and stable lattice structure, making migration within this phase more challenging. The disparity in lattice integrity between CsCu₂I₃ and CuI is pivotal in shaping the ion transport dynamics and the device’s electrical behavior. This difference also indirectly confirms that the presence of the CuI layer further limits current increases.

Voiceprint recognition based on MHDM

Leveraging its frequency-amplitude dependence, the MHDM enhances speaker-specific acoustic features by nonlinearly modulating conductance in the low-voltage regime, serving as a preprocessing layer for voiceprint recognition. As illustrated in Fig. 4a, normalized speech signals are converted into voltage pulses and applied to the device. The resulting dynamic current responses are then acquired and fed into a one-dimensional convolutional neural network (1D-CNN) for subsequent processing. The shallow layers of this network extract and consolidate the acoustic features already refined by the MHDM^49,50, demonstrating a hardware-software co-optimization strategy. This design reduces the computational overhead of the software backend, enabling efficient voiceprint encoding and final recognition model training using labeled data.

Fig. 4. Voiceprint recognition process based on the MHDM.

a Architecture of the proposed voiceprint preprocessing and recognition system. b Original voice waveforms recorded from four speakers (two males and two females). c Voiceprint conversion and output process. d Current outputs from the 1^st and 50^th operational cycles under identical voice inputs demonstrate stable feature extraction. e Comparison of voiceprint recognition accuracy for original signals, signals processed by the MHDM, and those processed by a single device over 50 cycles. f Confusion matrix of speaker classification based on MHDM-processed signals, confirming precise identity discrimination across four speakers.

To evaluate the proposed system, voiceprint recognition was performed using the spoken phrase “你好“ (the Chinese word of “hello”) recorded from four distinct individuals, including two males and two females (Fig. 4b). The speech signals were converted into voltage pulses and applied to the device, yielding a total of 4198 current responses that constituted the dataset (Fig. 4c). Notably, the diversity of this dataset was intrinsically enhanced by device-to-device and cycle-to-cycle variations, which introduced subtle but measurable differences in the current responses (Supplementary Fig. 23). These variations enriched the training data by exposing the model to a broader range of device-specific signatures, thereby improving its ability to generalize to real-world scenarios (Fig. 4d)^51,52. Based on measured parameters, our MHDM device exhibits an equivalent average power of approximately 0.075 to 11.4 μW, which is comparable to or lower than that of low-power speech recognition processors^53,54. In contrast, conventional digital implementations typically consume six orders of magnitude more energy per operation than the 0.02 to 2.85 nJ per pulse of our device⁵⁵. This energy-efficient operation is achieved while maintaining high fidelity in capturing individual vocal characteristics. The recognition system successfully mapped the resulting voiceprint vectors to specific speaker identities, achieving an accuracy of 99.3% when tested on signals refined by the MHDM (Fig. 4e). The corresponding confusion matrix (Fig. 4f) further confirms the precision of the system, establishing it as a viable candidate for secure authentication tasks.

To further assess the reliability and stability of this hardware-based preprocessing approach, repeated measurements were conducted over 50 operational cycles using identical voice samples. The signals processed by the device across these cycles were subsequently fed into the recognition backend, which maintained a consistent accuracy of 98.2% (Fig. 4e). This result confirms the excellent reproducibility of the device’s feature refinement capability. Notably, this performance is comparable to the 99.6% accuracy obtained from the original acoustic data before device processing, demonstrating the high-fidelity preservation of features by the proposed hardware-level approach. Environmental robustness was further evaluated by exposing the device to varying humidity levels during signal acquisition. Stable signal output was maintained across high-humidity conditions, and subsequent recognition tests yielded accuracy levels comparable to those obtained under standard operation (Supplementary Fig. 24). To further contextualize the device’s performance, it is compared with representative volatile memristors reported for temporal processing applications (Supplementary Table 1). While devices such as threshold switches and diffusive memristors exhibit decay times comparable to those of the proposed device (μs ~ ms), they suffer from thermal instability or filament stochasticity, which undermine their reliability for sustained speech-processing tasks^56–60. In contrast, our MHDM achieves fast decay without compromising stability. This unique combination positions our device as a distinctive solution for hardware-based voiceprint recognition.

Robust voiceprint recognition in noisy environments

Subsequently, to verify the robustness of the MHDM for voiceprint recognition, the system was trained on voice datasets containing more complex speech from different speakers, with multiple ambient noises. The speech signals were collected from experimenters using mobile phones in various environments, as shown in Fig. 5a. The quiet environment and four ambient noises of rain, machine, busy street, and echo were also mixed in these signals, respectively (Fig. 5b and Supplementary Note 2). These datasets were simultaneously acquired from 6 speakers and ten items for each. Supplementary Table 2 displays the temporal acoustic signals of all six speakers uttering “What’s your name” in a quiet environment. To expand the dataset and increase variability, we collected 10 additional speech samples, comprising 2 in Chinese and 8 in English (Supplementary Table 3). These samples were recorded under five distinct background noise conditions in real-world scenarios. Ultimately, 300 speech samples were gathered, providing a comprehensive dataset for subsequent signal processing and analysis. Based on variations in device performance, 1500 datasets were randomly selected from the collected samples for training and testing. The datasets were split at a 4:1 ratio, with 1200 samples used for training and 300 for testing. This approach allows for a balanced evaluation of the model’s performance, accounting for both device variability and the diversity of the collected speech samples.

Fig. 5. Self-filtering noise reduction and voiceprint recognition based on the MHDM.

a Schematic illustration of the voiceprint recognition system processing speech signals contaminated with various real-world background noises. b Waveforms of the phrase “What’s your name” recorded from a single speaker under five different acoustic conditions. c SNR of the speech signal with machine noise before and after MHDM processing. d SNR comparison for speech signals under four different noise conditions before and after MHDM processing. Data are presented as mean ± SD. e Voiceprint recognition accuracy under quiet conditions and MHDM-filtered noisy conditions. f The confusion matrix of speaker classification based on MHDM-processed speech signals under noisy conditions. g Voiceprint recognition accuracy for noisy speech signals before and after MHDM processing, evaluated on both the self-collected dataset and the public VoxCeleb dataset.

Figure 5c illustrates the performance of our hardware system for processing the voice signal of “What’s your name” in the presence of machine-generated background noise. Our device effectively processed speech signals, significantly improving the SNR from 10.5 to 15.8 dB in the output (Supplementary Note 2). The SNR enhancement highlights the device’s capability to suppress high-amplitude voltage signals under continuous pulse stimulation, which is crucial for noise resilience in practical applications. The noise-reduction performance for different speakers is illustrated in Supplementary Table 4, indicating that our device demonstrates strong generalization, effectively reducing noise across speakers. Furthermore, the device performs better in identifying speakers’ identities, indicating its robustness and adaptability in diverse acoustic environments. This result underscores the potential of our device for practical applications in noise-prone conditions while maintaining accurate speaker recognition^7,61.

Subsequently, we systematically evaluated the performance of the proposed system in processing speech signals under diverse background noise conditions. As summarized in Fig. 5d and Supplementary Table 5, the device consistently increases the average SNR by over 20% after signal processing across all noise scenarios, demonstrating effective and reliable noise suppression across varying acoustic environments. This demonstrates the MHDM system’s ability to adapt to diverse noise profiles while preserving the clarity of the original speech signal, further validating its potential for real-world applications. In contrast, traditional memristor-based devices used in speech signal processing have primarily focused on tasks such as speech classification, often relying on pre-existing databases. These systems are frequently limited by their inability to manage noisy environments or to distinguish between speakers. A comparative analysis of these parameters is presented in Supplementary Table 6, which outlines the advantages and limitations of each system. These results demonstrate that our MHDM system achieves robust noise-adaptive speech enhancement, surpassing conventional memristor-based approaches. Furthermore, classification using the MHDM achieved voiceprint recognition accuracy of 93.2% for speech signals from noisy environments, reaching a level comparable to that obtained under quiet conditions with an accuracy of 95.9% (Fig. 5e). The confusion matrix in Fig. 5f further confirms the system’s robust performance in classifying speaker identities within complex acoustic scenes.

To assess the generalizability of our approach, we extended the evaluation to the VoxCeleb dataset, a large-scale, open-source collection of real-world outdoor speech⁶². Details of the data preprocessing and device mapping protocols are provided in Supplementary Notes 3 and 4. As shown in Fig. 5g, for the self-collected dataset, the recognition accuracy improved from 90.6% for original noisy signals to 93.2% after processing through the device. Similarly, for the VoxCeleb dataset, conventional software-based classification of the raw speech data yielded an accuracy of 95.3%, whereas processing the same signals via the device’s nonlinear mapping increased recognition accuracy to 98.7%. These results demonstrate the device’s noise-suppression capability, thereby enhancing voiceprint recognition accuracy across diverse environmental conditions.

Discussion

In summary, this study presents an MHDM with a gradient-distributed functional layer, offering an innovative hardware platform for effective voiceprint recognition in challenging acoustic environments. The gradient-tuned halide composition allows for precise modulation of the Schottky barrier and ion migration, facilitating adaptive signal processing. With a signal attenuation time of less than 10 μs, the memristor efficiently encodes high-frequency speech components. At the same time, its intrinsic nonlinearity enables the extraction of discriminative features even with large-amplitude inputs. Impressively, the system consistently achieves an SNR improvement of over 20% across various noise scenarios, resulting in a recognition accuracy of 99.3% in quiet environments and 93.2% retention in noisy settings. These findings underscore the MHDM’s potential for secure, edge-deployable authentication and real-time speech processing applications.

Methods

Device fabrication

The MHDM consisted of ITO/Cs-Cu-I/Au. Glass substrates with pre-patterned ITO electrodes were cleaned by sequentially sonicated in acetone, ethanol, and deionized water for 10 min each, then baked in an oven at 120 °C for 30 min, followed by a UV treatment for 10 min. The Cs-Cu-I film with a thickness of 80 nm was deposited by thermal evaporation in vacuum (≈10⁻⁴ Pa), in which the heat source was a mixture of CsI and CuI with a molar ratio of 1:2. The thickness of the entire film during deposition is controlled by the evaporation rate and thickness of CsI. The device had a working area of 100 µm × 100 µm. Au top electrodes were finally thermally evaporated onto the Cs-Cu-I layer through patterned shadow masks in a vacuum (≈10⁻⁴ Pa).

Characterization

The device was measured using an electrical measurement system consisting of a Keithley 2636B semiconductor parameter analyzer, a PDA FS-Pro parameter analyzer, and a Lakeshore CRX-6.5 K cryogenic probe station for cryogenic measurements. The surface morphology of Cs-Cu-I film was investigated using atomic force microscopy (AFM, Bruker’s Dimension Icon Atomic Force Microscopy) operated in the tapping mode. SEM micrographs were obtained using field-emission scanning electron microscopy (Hitachi S4800-SEM). The film crystal phase was confirmed through an X-ray diffractometer (D8 Advance A25, Bruker). All XPS data were acquired using a PHI 5000 VersaProbe (ULVAC-PHI) system. All ToF-SIMS depth profiles were acquired using a PHI TRIFT V nano TOF (ULVAC-PHI) system with a dual beam slice-and-view scheme. The cross-sectional analysis of the Cs-Cu-I memristor was conducted using focused ion beam (FIB) systems and HRTEM technology, equipped with EDS.

Speech data acquisition and preprocessing

The speech corpus used in this study comprised two distinct datasets for device-level evaluation. For the clean speech recognition task, utterances of the Chinese word “你好“ were collected from four individual speakers. Each utterance was digitally sampled at a rate of 18,000 points, yielding 18,000 discrete samples per utterance. These time-domain amplitude values were subsequently normalized and linearly mapped to the voltage range of 0–0.3 V, corresponding to the low-bias operational regime of the memristor. The resulting voltage sequence was then sequentially applied to the device as input stimuli at a rate of one data point every 0.25 ms. The output current from the device was recorded and subsequently used for feature extraction and speaker classification.

For the noisy speech recognition task, a more diverse dataset was employed, consisting of ten distinct sentence-type utterances collected from six individual speakers. Each utterance was sampled at 60,000 points, resulting in 60,000 discrete samples per utterance. The amplitude values were normalized and linearly mapped to an extended voltage range of 0–0.5 V to leverage the device’s enhanced nonlinearity at higher bias. The voltage sequence was applied to the memristor at an input rate of 1 data point every 0.25 ms, identical to that of the clean speech task. The output current signals were recorded for subsequent processing and speaker identification.

Neural network architecture for MHDM-refined voiceprint recognition

A five-layer 1D-CNN was constructed for speaker recognition tasks. The network accepts single-channel raw audio waveforms of 18000 sampling points as input, corresponding to audio segments extracted from the voice dataset. Each audio segment was preprocessed using zero-mean unit-variance standardization (Z-score normalization) applied independently to each of the 18,000 temporal features. The network architecture consists of five convolutional blocks followed by three fully connected layers. The first convolutional layer employs a large kernel size (kernel size = 80, stride = 4) to capture low-frequency acoustic features from the raw waveform, producing 64 feature maps. This is followed by batch normalization, ReLU activation, max-pooling (pool size = 4), and dropout (rate = 0.1). The subsequent four convolutional layers progressively extract higher-level features with increasing channel dimensions (128, 256, 512, 512) using 3 × 1 kernels, each followed by batch normalization, ReLU activation, and max-pooling layers (pool sizes: 4, 4, 2, and adaptive global average pooling). Dropout rates gradually increase from 0.1 to 0.4 across the convolutional blocks to prevent overfitting. After global average pooling, the 512-dimensional feature vector is passed through three fully connected layers with 256, 128, and finally 6 output neurons corresponding to the six speaker classes. Each fully connected layer incorporates batch normalization and dropout regularization (rate = 0.5), with ReLU activations applied to the first two layers. The final layer produces class probability distributions via the softmax function.

Author contributions

H.S. and J.Y.M. contributed equally to the work. H.S. and J.Y.M. conceived the idea and proposed the research. H.S. designed and conducted the experiments. J.Y.M. and W.Y. performed the characterization of materials and electrical measurements. R.H.W. was responsible for the data collection, processing, and recognition. X.H., B.X.L., J.T.S., Y.M., L.G., and W.L. provided mechanism analysis and discussion. H.F.L., L.H.X., and W.H. supervised the project. H.F.L. and J.C.H. revised the manuscript and provided additional suggestions. All authors contributed to the manuscript revision.

Peer review

Peer review information

Nature Communications thanks Jianshi Tang and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

H.S. discloses support from the National Natural Science Foundation of China [grant number 12204248]. H.F.L. discloses support from the National Natural Science Foundation of China [grant number 62471251], the Basic Research Program of Jiangsu [grant number BK20240033], and the National Key Research and Development Program of China (grant number 2021YFA0717900). W.H. discloses support from the National Natural Science Foundation of China [grant number 62288102], the Basic Research Program of Jiangsu [grant number BK20243057]. Y.M. discloses support from the National Natural Science Foundation of China [grant number 62574077]. J.Y.M. discloses support from the Postgraduate Research & Practice Innovation Program of Jiangsu Province [grant number SJCX21_0252].

Data availability

All data that support the findings of this study are included in the published article and its Supplementary Information. Additional data is available from the corresponding author on request. The VoxCeleb dataset used in this study was accessed via https://www.robots.ox.ac.uk/~vgg/data/voxceleb/. Source data are provided with this paper.

Code availability

The code used for the software simulation in this study is available from the corresponding authors upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-74047-3.