Paper deep dive
Efficient Audio-Visual Speech Enhancement via Reinforcement Learning-Based Neural Architecture Search
Khubaib Ahmed, Ahsan Adeel, Nasir Saleem, Kia Dashtipour, Ahsan Ulhaq, Amir Hussain
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 10/12/2026, 6:22:07 AM
Summary
This paper introduces a Reinforcement Learning-based Neural Architecture Search (RL-NAS) framework for Audio-Visual Speech Enhancement (AVSE). The framework uses Proximal Policy Optimization (PPO) with constraint-aware action masking to discover compact architectures that balance speech enhancement quality (SI-SNR) and model size. The resulting model has 2.9 million parameters (a 9-fold reduction from the 22M baseline) with only a 0.7 dB drop in SI-SNR improvement, achieving a 4.3-fold reduction in real-time factor on mobile CPUs. The search revealed that shallower, wider temporal convolutional blocks and aggressive visual pathway pruning are effective for efficiency.
Entities (8)
Relation Signals (7)
RL-NAS â produces â 2.9 million parameters
confidence 98% · identifies a model with 2.9 million parameters, compared with 22 million in the baseline.
RL-NAS â optimizesfor â Si-SNR
confidence 95% · a novel reward function jointly optimises speech enhancement performance, measured by SI-SNR improvement, and model compactness.
RL-NAS â optimizesfor â Model Size
confidence 95% · a novel reward function jointly optimises speech enhancement performance, measured by SI-SNR improvement, and model compactness.
RL-NAS â uses â PPO
confidence 95% · Using proximal policy optimisation with constraint-aware action masking, the framework evaluates only 35 candidate architectures...
RL-NAS â outperforms â Random Search
confidence 93% · Compared with random search, RL-NAS achieves 92% higher search efficiency.
Temporal Convolutional Network â iscomponentof â AVSE
confidence 92% · Architectural analysis reveals that shallower but wider temporal convolutional blocks... can maintain enhancement performance...
RL-NAS â reduces â Real-Time Factor
confidence 90% · Evaluation on commercial mobile CPUs demonstrates a 4.3-fold reduction in real-time factor...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Audio-visual speech enhancement (AVSE) improves speech intelligibility in noisy environments but often relies on computationally expensive models unsuitable for resource-constrained edge devices. This paper introduces a reinforcement learning-driven neural architecture search (RL-NAS) framework to discover compact, efficient AVSE architectures. The search space preserves temporal alignment between audio and visual streams, while a novel reward function jointly optimises speech enhancement performance, measured by SI-SNR improvement, and model compactness. Using proximal policy optimisation with constraint-aware action masking, the framework evaluates only 35 candidate architectures and identifies a model with 2.9 million parameters, compared with 22 million in the baseline. The proposed architecture achieves a 14.5 dB SI-SNR improvement, only 0.7 dB below the baseline, while reducing model size by approximately nine-fold. Evaluation on commercial mobile CPUs demonstrates a 4.3-fold reduction in real-time factor, supporting practical on-device deployment. Compared with random search, RL-NAS achieves 92% higher search efficiency. Architectural analysis reveals that shallower but wider temporal convolutional blocks and aggressive visual pathway pruning can maintain enhancement performance while substantially reducing computational requirements. These findings demonstrate the potential of reinforcement learning-guided architecture search for developing lightweight, efficient and deployable multimodal speech enhancement systems.
Tags
Links
- Source: https://arxiv.org/abs/2610.12796v1
- Canonical: https://arxiv.org/abs/2610.12796v1
Trouble viewing inline? Open PDF directly â
Full Text
47,745 characters extracted from source content.
Expand or collapse full text
Efficient Audio-Visual Speech Enhancement via Reinforcement Learning-Based Neural Architecture Search Khubaib Ahmed Ahsan Adeel Nasir Saleem Kia Dashtipour Ahsan Ulhaq Amir Hussain Abstract Audio-visual speech enhancement (AVSE) has emerged as a critical technology for enabling reliable human-computer interaction in challenging acoustic environments such as crowded cafĂ©s and moving vehicles. However, state-of-the-art AVSE models frequently exceed 22 million parameters, making them unsuitable for real-time deployment on resource-constrained edge devices such as smartphones, wearables, and embedded systems. To address this fundamental deployment challenge, we introduce a reinforcement learning-driven neural architecture search (RL-NAS) framework that automatically discovers compact yet high-performance AVSE architectures. Our approach constrains the design space to respect the temporal alignment requirements between audio and visual streams while carefully controlling parameter proliferation through a novel reward function that jointly optimizes speech-quality improvements (measured by SI-SNR gains) and model compactness. By leveraging proximal-policy optimization with constraint-aware action masking, our agent efficiently evaluates only 35 candidate architectures before converging on a 2.9-million-parameter model achieving 14.5 dB SI-SNR improvementâmerely 0.7 dB below the 22-million-parameter baseline, yet with a nine-fold compression ratio. When deployed on commercial mobile CPUs, this lightweight network achieves a 4.3Ă reduction in real-time factor, demonstrating practical viability for on-device speech enhancement. In comparison to uninformed random search, our RL-driven methodology achieves 92% superior search efficiency while revealing crucial architectural principles, including the preference for shallower but wider temporal-convolutional blocks and the tolerance of aggressive visual-pathway pruning. These discoveries establish new design principles for efficient multimodal speech systems and demonstrate that hardware-aware NAS, guided by reinforcement learning, can successfully deliver lightweight AVSE models without sacrificing perceptual quality. keywordsAudio-visual speech enhancement, Neural architecture search, Model compression, Reinforcement learning, Efficient speech processing â address: 1University of Stirling, UK, 2Edinburgh Napier University, UK, 3University of Technology and Applied Sciences, Omanâ email: khubaib.ahmed@stir.ac.uk1, ahsan.adeel1@stir.ac.uk1, N.Saleem@napier.ac.uk2, K.Dashtipour@napier.ac.uk2, ahsan.ulhaq@utas.edu.om3, A.Hussain@napier.ac.uk2 1 Introduction Audio-visual speech enhancement represents a paradigm shift in robust speech processing, leveraging the complementary information contained in visual lip movements to augment noisy acoustic signals [1, 2, 3]. This multimodal approach consistently outperforms audio-only alternatives, particularly in scenarios characterized by severe acoustic degradation. The field of lip reading has advanced considerably in recent years, with methods like LipNet [4] and broader visual speech recognition frameworks [5] providing strong feature extraction foundations. Recent advances leveraging temporal convolutional networks have substantially improved enhancement quality, with architectures such as AVSEGAN [6] and Conv-TasNet variants [7] achieving remarkable performance metrics. General speech enhancement techniques have also evolved significantly, with supervised approaches [8], speech separation methods [9, 10], and enhancement frameworks [11, 12] all contributing to the knowledge base. However, these state-of-the-art systems carry a prohibitive computational burden, with model sizes frequently exceeding 22 million learnable parameters. This substantial parameter count creates significant barriers to practical deployment, preventing the adoption of AVSE technology on smartphones, hearing aids, and other edge devices where computational resources are inherently limited. Neural Architecture Search offers a promising avenue for addressing this efficiency gap. By automating the discovery of optimal network topologies, NAS can explore vast design spaces and identify architectures that balance performance and efficiency far more systematically than manual engineering approaches. Nevertheless, applying NAS to multimodal speech enhancement presents unique and substantial challenges. The interdependencies between audio and visual processing pathways create an enormous potential search space. Moreover, the requirement to maintain temporal synchronization between the audio and visual streams introduces architectural constraints that must be carefully enforced. Additionally, the training cost associated with evaluating candidate architectures becomes compounded when multiple sensor modalities are involved. To overcome these challenges, we propose a comprehensive RL-based NAS framework specifically designed for audio-visual speech enhancement. Our framework incorporates three fundamental innovations: (1) a carefully designed constrained search space that explicitly preserves temporal alignment between modalities while maintaining architectural validity, (2) a sophisticated parameter-performance balanced reward function that elegantly trades off enhancement quality against model size, and (3) a progressive shrinking strategy that efficiently guides the search from high-performance baseline architectures toward increasingly compact solutions. Through extensive experimentation, we demonstrate that our approach discovers a compact 2.9-million-parameter model that retains 95.4% of the baseline performance while achieving a nine-fold reduction in model size, enabling practical deployment on edge devices. Our RL-NAS methodology also reveals important architectural insights that challenge conventional wisdom about multimodal network design, including the surprising finding that visual information can be represented in dramatically reduced dimensional spaces without substantial performance degradation. 2 Related Work 2.1 Neural Architecture Search Foundations Neural Architecture Search has fundamentally transformed the landscape of automated machine learning by enabling the systematic discovery of optimal network architectures without extensive manual engineering. The seminal work by Zoph and Le introduced reinforcement learning as a powerful tool for this purpose [13], where an RL agent learned to sequentially construct network architectures by predicting the configuration of successive layers. Evolutionary approaches to NAS have also proven effective, with regularized evolution methods [14] demonstrating competitive performance without the stability issues associated with early RL-based approaches. However, the computational requirements of these initial methods, which necessitated thousands of GPU-days for complete architecture discovery, motivated substantial subsequent research into more efficient methodologies. The introduction of differentiable NAS (DARTS) represented a pivotal advancement by reformulating architecture search as a continuous optimization problem amenable to gradient-based methods [15]. This perspective shift dramatically reduced the computational cost of architecture search. Subsequent innovations focused on further improving search efficiency through parameter sharing strategies [16], which enable the RL agent to leverage knowledge accumulated during the evaluation of previous architectures. Hardware-aware NAS methodologies like ProxylessNAS [17] have emerged to directly optimize for deployment targets. More recent approaches employ predictor-based methods that learn to estimate architecture performance without requiring full model training [18], dramatically accelerating the search process. Training-free NAS methods [19] represent an even more extreme efficiency frontier. Contemporary research increasingly emphasizes multi-objective optimization formulations that simultaneously consider multiple competing objectives such as model accuracy, computational latency, and memory consumption [20], particularly important for edge deployment scenarios where computational resources are severely constrained. Recent advances in generative NAS [21] and LLM-informed design principle transfer [22] suggest promising new directions for architecture discovery. 2.2 NAS Applications in Speech and Audio Processing Speech and audio processing tasks present distinctive challenges for neural architecture search due to their inherent temporal dynamics and stringent noise robustness requirements. ST-NAS pioneered the application of differentiable architecture search to end-to-end automatic speech recognition [23], demonstrating that NAS-discovered architectures could match or exceed the performance of manually designed systems. Parallel work by Hu et al. introduced NAS-TDNN, which automated the architecture search process for time-delay neural networks and achieved remarkable 96% model compression while preserving accuracy [24]. These successes motivated intensive investigation into optimal architectures for speech enhancement specifically. Temporal convolutional networks emerged as particularly well-suited substrates for this task, offering an attractive balance between effective long-range temporal modeling and computational efficiency [7]. The demonstrated success of TCN-based architectures in speech enhancement applications has made them the foundation for numerous subsequent competitive systems. 2.3 Optimal Temporal Convolutional Network Design Systematic investigation through neural architecture search has revealed consistent and reproducible patterns in optimal TCN design for speech processing tasks. Research has demonstrated that a subtle but important depth-width tradeoff exists: shallower networks with higher expansion ratios (typically 1.5 to 2.0) consistently outperform deeply nested narrower configurations, contrary to intuitions inherited from computer vision architectures [25]. Temporal convolutional networks have emerged as a powerful paradigm for sequence modeling tasks [26, 27], with particular strength in capturing long-range dependencies through dilated convolutions [28]. Furthermore, the dilation schedule employed in dilated convolutional blocks significantly influences performance. Analysis across multiple papers has shown that linear dilation stride patterns systematically surpass exponential dilation strategies, particularly when the network must maintain robustness to environmental noise [29]. An important architectural principle involves parameter redistribution across layers: shifting learnable parameters from later processing stages toward earlier layers creates more robust feature representations and enhances overall noise tolerance [6]. These principles informed the design of more recent competitive architectures such as MSGLN, which integrates global self-attention mechanisms with local convolutional operations to efficiently model long-range temporal dependencies while preserving computational efficiency [25]. 2.4 Multimodal and Efficient Neural Architecture Search The extension of neural architecture search to multimodal learning scenarios remains an active and challenging research frontier. The coordination of cross-modal fusion strategies and the balancing of architectural investments across different modalities present substantial technical challenges. Recent work has begun to address these challenges through innovative approaches such as architecture embedding vectors that enable knowledge transfer across diverse vision and audio tasks [30]. Multimodal learning theory and practice [31, 32, 33] provide crucial foundations for understanding how to integrate multiple information streams effectively. Audio-visual research has established several crucial principles for multimodal architecture design. Notably, visual processing pathways demonstrate substantially higher tolerance for aggressive parameter reduction compared to audio streams, with studies demonstrating that 50% to 80% reductions in visual pathway complexity frequently yield minimal performance degradation [1]. This asymmetry arises because visual lip movement information is relatively simple geometrically compared to the complex spectral and temporal patterns in degraded audio. Fusion strategy selection also critically impacts efficiency: additive fusion mechanisms typically outperform concatenation-based approaches while simultaneously reducing the overall parameter count by approximately 85% [29]. Encouraging preliminary evidence suggests that sub-3-million-parameter models can maintain performance within 1 dB SI-SNR of 20+ million parameter baselines [6], motivating the search for even more aggressive compression strategies. Model compression and efficiency optimization remain essential for practical deployment. Pruning-based approaches [34, 35] and quantization methods [36, 37] have demonstrated substantial parameter reductions. The success of efficient mobile architectures like MobileNets [38, 39] and MobileBERT [40] has established design principles for resource-constrained inference. Emerging approaches such as binarized neural architecture search and normalizing flow-based design methodologies [41] represent promising directions for achieving further compression in low-parameter regimes. 2.5 Generative Approaches to Speech Enhancement Recent work has explored generative adversarial networks (GANs) [42] as an alternative paradigm for speech enhancement. AVSEGAN represents an important application of GAN-based learning to audio-visual enhancement, leveraging adversarial training dynamics. Beyond speech-specific applications, generative approaches like pix2pix [43] have demonstrated effectiveness in image-to-image translation tasks that share similar structural challenges with audio-visual enhancement. However, GANs introduce additional complexity and training instability compared to supervised approaches, making them less amenable to efficient hardware deployment. Our architecture search framework sidesteps these challenges by focusing on discriminative learning with explicit efficiency constraints. 2.6 Research Gaps and Our Contribution Despite substantial progress in neural architecture search and speech enhancement, important gaps persist in the current literature. Current multimodal speech enhancement systems typically lack generalization mechanisms to handle variations across different noise types and signal-to-noise ratio conditions [7]. Most systems operate under the assumption of fixed acoustic conditions rather than adapting dynamically to real-time environmental changes. The intersection of quantization-aware architecture search with speech enhancement remains substantially underexplored, despite the critical importance of low-bit representations for edge deployment. Hardware-aware NAS for speech tasks remains an underexplored area, with most efficiency-oriented research focused on vision and NLP domains [44]. Our work directly addresses these gaps through several key contributions. First, we introduce constrained search spaces that explicitly preserve temporal alignment requirements while dramatically reducing the effective search complexity. Second, our parameter-performance balanced reward function elegantly trades off multiple competing objectives through a principled multi-objective optimization approach. Third, our framework incorporates hardware-aware efficiency considerations from initial design through final deployment. Finally, our comprehensive analysis reveals reproducible architectural patterns and design principles that can guide future multimodal speech-enhancement research, providing a foundation for continued progress in this critical domain. 3 RL-Based NAS Methodology Our reinforcement learning-based neural architecture search framework implements a constrained Markov Decision Process that systematically optimizes audio-visual architectures while rigorously preserving the essential temporal synchronization between modalities. The methodology extends established hardware-aware NAS principles [20] with novel modality-specific innovations carefully tailored for the unique requirements of speech enhancement. 3.1 Constrained Search Space Design The search space design process begins with domain insights gleaned from extensive analysis of temporal convolutional network optimization studies. Rather than exploring the full combinatorial explosion of possibilities, we incorporate domain knowledge to enforce three categories of constraints that dramatically reduce the effective search complexity while ensuring that all discovered architectures remain architecturally viable. The temporal consistency constraint ensures that audio and visual processing pathways maintain perfect temporal alignment throughout the network: Taudio=TvisualT_audio=T_visual. This requirement is enforced through the exclusive use of dimension-preserving operations (such as pointwise convolutions and residual connections) that do not alter the temporal resolution of intermediate feature maps. The channel compatibility constraint maintains the mathematical validity of skip connections and fusion operations by enforcing Bâ€EB†E, where B denotes bottleneck channels and E represents encoder channels. The compute bounds constraint prevents runaway parameter growth by limiting RĂBâ€24RĂ B†24, where R is the number of TCN repeats and B is the block dimension, thereby preventing memory exhaustion during training and deployment. Table 1 presents the complete search space configuration. This carefully constrained design reduces the potential search space from âĄ(1014)O(10^14) unconstrained possibilities to merely 324 valid architectural configurations, representing a 99.9997% reduction in search complexity while maintaining sufficient diversity to discover novel and high-performing architectures. 3.2 Reinforcement Learning Formulation We model the architecture search process as a finite-horizon Markov Decision Process with carefully designed state, action, and reward components. The state representation stâs_t encodes both architectural parameters and accumulated performance metrics: st=[E256,B256,R,X,HB,Params22âM,SI-SNR15.2]s_t= [ E256, B256,R,X, HB, Params22M, SI-SNR15.2 ] (1) Each dimension undergoes min-max normalization to the [0,1] interval using the established embedding strategy from architecture transfer learning, augmented with performance-aware components that explicitly account for the current quality of speech enhancement achieved by candidate architectures. This design ensures that the RL agent maintains awareness of both structural properties and empirical performance throughout the search trajectory. The action space A comprises 12 discrete operations: increase and decrease operations for six architectural dimensions. Critically, we implement constraint-aware action masking that restricts the agent to valid transitions: valid=aâ|satisfiesâ(st,a,)A_valid=\a \,|\,satisfies(s_t,a,C)\ (2) This masking mechanism eliminates invalid architecture configurations from consideration and substantially accelerates convergence compared to penalty-only approaches, achieving approximately 40% faster convergence by preventing the RL agent from wasting samples on invalid architectures. 3.2.1 Multi-Objective Reward Function Design The reward function represents the core mechanism through which we guide the RL agent toward desirable architectural trade-offs. Rather than optimizing a single scalar objective, we formulate a multi-objective reward function that elegantly balances three competing goals: maximizing speech enhancement quality, minimizing model size, and enforcing architectural constraints. The design reflects a careful balance between these objectives: RâĄ(st)=αâÎâSI-SNRBaselineâFidelityâÎČâmaxâĄ(0,ParamsTargetâ1)âEfficiencyâÎłâViolationâConstraintR(s_t)= α -SNRBaseline_Fidelity- ÎČ (0, ParamsTarget-1 )_Efficiency- Îł 1_Violation_Constraint (3) The fidelity component encourages the agent to discover architectures that maintain high speech enhancement quality, normalized by the baseline performance to provide stable learning signals across different noise conditions. The efficiency component penalizes excess parameter counts only when they exceed the target budget of 2.9 million parameters, avoiding unnecessary penalties for early-stage architectures that legitimately require larger size. The constraint component imposes a severe penalty when any architectural constraint is violated, effectively preventing the exploration of invalid configurations. The coefficients (α=1.0α=1.0, ÎČ=0.7ÎČ=0.7, Îł=10Îł=10, target=2.9M) were determined through systematic multi-objective Pareto analysis, ensuring that discovered architectures form the true Pareto frontier of efficiency-performance trade-offs rather than arbitrary compromises. The specific value of ÎČ=0.7ÎČ=0.7 reflects our empirical finding that a moderate efficiency penalty encourages compression without driving the search toward undersized, ineffective models. 3.3 Policy Network Architecture The RL agent employs a 2-layer LSTM with 256 hidden units per layer, chosen specifically for its documented ability to model the sequential dependencies inherent in architecture transformation sequences. This architecture reflects established best practices in deep reinforcement learning for discrete action spaces. The network maintains two distinct output heads: an actor head that generates the action probability distribution through softmax normalization, and a critic head that estimates the current state value for advantage calculation: Action distribution :ÏâĄ(a|s)=softmaxâ(Linearactorâ(t)) :Ï(a|s)=softmax(Linear_actor(h_t)) State value :vâĄ(s)=Linearcriticâ(t) :v(s)=Linear_critic(h_t) where th_t denotes the final hidden state from the recurrent layers. A key initialization technique involves seeding the LSTM with learned embeddings of the baseline 22-million-parameter architecture, effectively giving the agent a warm start that accelerates exploration toward high-performing regions of the search space. State Vector LSTM Layer 1 (256 units) LSTM Layer 2 (256 units) Actor Head (Action logits) Critic Head (Value estimate) Action Masking ata_tvtv_tRL Agent Architecture Figure 1: Policy network architecture: LSTM processor with specialized heads for action selection and value estimation. Action masking enforces architectural constraints. 3.4 Training with Proximal Policy Optimization The RL agent undergoes training through Proximal Policy Optimization [45], selected for its proven sample efficiency and training stability characteristics compared to alternative policy gradient methods [46, 47]. PPO employs a clipped objective function that prevents excessively large policy updates which could destabilize learning: LCLIPâ(Ξ)=tâ[minâĄ(rtâ(Ξ)âA^t,clipâ(rtâ(Ξ),1âÏ”,1+Ï”)âA^t)]L^CLIP(Ξ)=E_t [ (r_t(Ξ) A_t,clip(r_t(Ξ),1-Δ,1+Δ) A_t ) ] (4) where rtâ(Ξ)=ÏΞâ(at|st)ÏΞoldâ(at|st)r_t(Ξ)= _Ξ(a_t|s_t) _ _old(a_t|s_t) quantifies the probability ratio between the current and previous policies, A^t A_t denotes the generalized advantage estimate, and Ï”=0.2Δ=0.2 controls the clipping region. This formulation ensures that policy updates remain bounded, preventing the kind of catastrophic forgetting that can plague RL algorithms during exploration. The training implementation incorporates several key design decisions. Exploration employs ϔΔ-greedy action selection with Ï”=0.3Δ=0.3, combined with our constraint-aware action masking to ensure all exploration remains within valid architecture space. The optimization algorithm uses AdamW with a learning rate of 3Ă10â43Ă 10^-4 and weight decay of 10â210^-2 to maintain stable gradient-based updates. We employ four parallel workers for simultaneous architecture evaluation, enabling efficient wall-clock training time. The generalized advantage estimation with λ=0.95λ=0.95 provides low-variance advantage estimates that stabilize value-based learning. 3.5 Progressive Shrinking Strategy Algorithm 1 implements our progressive shrinking strategy, which orchestrates a curriculum learning approach that progressively compresses the baseline architecture toward the target parameter budget. This strategy incorporates three complementary mechanisms. First, curriculum learning begins the search from the high-performance 22-million-parameter baseline architecture rather than random initialization, providing a warm start that dramatically accelerates convergence. Second, step-wise dimension reduction modifies only a single architectural dimension per optimization step, maintaining architectural viability and enabling the agent to thoroughly explore the consequences of each modification. Third, fidelity extrapolation estimates the full-training SI-SNR performance from intermediate 10-epoch results using linear regression fit to historical data, dramatically reducing the computational cost of architecture evaluation. This systematic approach achieves 92% search efficiency compared to random exploration, requiring only 35 total architecture evaluations to identify high-performing configurations that would require thousands of random trials to discover. 3.6 Computational Complexity Analysis The total search cost comprises 35 episodes multiplied by 5 evolutionary steps per episode multiplied by 10 epochs per evaluation, yielding 1,750 epoch-equivalents of total computation. This represents a substantial reduction compared to alternative NAS methodologies: differentiable NAS typically requires 3,000+ GPU-hours, evolutionary NAS demands 2,500+ epochs, and exhaustive random search would necessitate 5,000+ epochs to achieve equivalent architectural coverage. Our constrained design space and progressive strategy yield approximately 3Ă speedup over generic NAS methods, making architecture search feasible for multimodal speech enhancement. Algorithm 1 Progressive Shrinking NAS with RL Agent 1: Initialize agent ÏΞ _Ξ with random weights 2: s0âs_0â Baseline architecture (22M params) 3: for episode = 1 to 35 do 4: Ïâ[]Ïâ[] Trajectory buffer 5: for step = 1 to 5 do 6: atâŒÏΞâ(st)a_t _Ξ(s_t) Sampled action 7: Apply action with constraints (Table 1) 8: Train model for 10 epochs 9: Evaluate SI-SNR, parameters 10: rtâRâĄ(st)r_tâ R(s_t) Compute reward 11: ÏâÏâȘ(st,at,rt)ÏâÏâȘ(s_t,a_t,r_t) 12: end for 13: Update Ξ using PPO with Ï 14: s0âs_0â Best architecture from Ï 15: end for Table 1: Constrained Search Space Configuration Component Parameters Options Constraint Encoder Channels 128, 192, 256 Eâ„128Eâ„ 128 Bottleneck Channels 128, 192, 256 Bâ€EB†E TCN Blocks per repeat 4,6,8 â Repeats 2,3,4 RĂBâ€24RĂ B†24 Expansion ratio 1.5,2.0,2.5 â Visual Output channels 128,256 Vâ„128Vâ„ 128 4 Experiments and Results 4.1 Experimental Setup Our evaluation framework emphasizes realistic audio-visual speech enhancement scenarios. We conduct experiments on the Lip Reading Sentences 3 (LRS3) dataset, which provides 438 hours of high-quality audio-visual speech recordings with synchronized audio and video streams [48]. For additional robustness analysis, we note that large-scale speech datasets like LibriSpeech [49] have become standard for evaluating speech enhancement and recognition systems. For noise, we employ the noise set from the Clarity Enhancement Challenge 2 (CEC2), which was specifically curated to reflect realistic hearing aid scenarios with signal-to-noise ratios ranging from â5-5 to 55 dB [50]. The evaluation metric is SI-SNR improvement in decibels, which measures the degree to which the network improves the signal-to-interference-plus-noise ratio. We also report the real-time factor on Intel i7-1185G7 mobile CPUs, which provides a practical measure of deployment feasibility. Speaker verification tasks on standard benchmarks like speaker recognition challenges [51] provide complementary evaluation perspectives for multimodal speech systems. The baseline model consists of a 22-million-parameter audio-visual speech enhancement network following the architecture described in previous work. 4.2 Architecture Evolution During Search Table 2 presents the progression of architectures discovered during the search process, showing the top trajectory from our five parallel search workers. The baseline 22-million-parameter architecture achieves a baseline SI-SNR improvement of 15.2 dB. As the RL agent progressively applies architectural modifications, we observe a carefully managed degradation in enhancement quality coupled with dramatic improvements in efficiency. After five optimization steps, the agent reduces the encoder channels from 256 to 192, achieving a 26.5% parameter reduction while maintaining nearly identical SI-SNR performance. Subsequent steps involve bottleneck reduction, TCN depth reduction, hidden dimension scaling, and finally aggressive visual pathway pruning. Each modification was made with careful attention to maintaining the viability of the resulting architecture and preserving adequate capacity for high-quality speech enhancement. Table 2: Architecture Progression Through Progressive Shrinking Search Step Params (M) SI-SNRi (dB) Î -SNR Reward Key Changes 0 22.0 15.2 0.00 0.00 Baseline 5 16.2 15.1 -0.1 0.72 E:256â 192 10 8.4 15.0 -0.2 1.25 B:256â 192 15 5.1 14.9 -0.3 1.87 R:4â 3 20 3.8 14.7 -0.5 2.01 X:8â 6 25 3.1 14.6 -0.6 2.13 H:2.5â 2.0 30 2.9 14.5 -0.7 2.31 V:256â 128 4.3 Discovered Optimal Architecture The neural architecture search process converged on a 2.9-million-parameter model that achieves excellent efficiency-performance balance. Compared to the baseline architecture, the discovered model incorporates several key modifications. The encoder channels are reduced from 256 to 192 dimensions, representing a 25% reduction in the feature representation capacity at the input stage. The bottleneck, which serves as the compression layer during processing, similarly reduces from 256 to 192 channels. The TCN component, which performs the core temporal modeling, employs only 3 repeats of 6-block configurations (43% reduction in temporal processing depth) with an expansion ratio of 2.0 in the hidden dimensions. Most dramatically, the visual processing pathway is aggressively pruned from 256 output channels to 128, representing a 50% reduction. This aggressive pruning of the visual pathway represents one of the most significant discoveries from the search process, as it suggests that lip movement information can be represented in dramatically lower-dimensional spaces than previously appreciated. 4.4 Performance Comparison Table 3 presents a comprehensive benchmarking comparison of our NAS-optimized model against multiple baselines and alternative approaches. Our 2.9-million-parameter model achieves 14.5 dB SI-SNR improvement, positioning it at 95.4% of the 15.2 dB performance achieved by the 22-million-parameter baseline. Critically, this performance is maintained while reducing model size by 87%. Compared to the handcrafted small model (6.1M parameters), our NAS-discovered architecture achieves 0.3 dB higher performance with 2.1Ă fewer parameters, demonstrating the value of systematic search over manual architecture design. The random search baseline, which achieved 13.9 dB SI-SNR with 2.4 million parameters, demonstrates that our RL-driven search achieves 0.6 dB higher performance at the same parameter budget. The real-time factor of 0.19 on mobile CPUs enables processing of 3-second audio files in approximately 570 milliseconds, making our model suitable for real-time interactive applications. Table 3: Performance Benchmarking on CEC2 Test Set Model Params (M) SI-SNRi (dB) RTF (CPU) Audio-only baseline 5.2 10.3 0.15 AVSE baseline 22.0 15.2 0.82 Handcrafted small 6.1 14.8 0.31 Random search best 2.4 13.9 0.22 NAS-optimized (ours) 2.9 14.5 0.19 Several important observations emerge from this benchmarking study. First, the efficiency-accuracy tradeoff demonstrates that our model achieves a near-optimal balance, delivering 95.4% of baseline enhancement quality with merely 13% of the parameters. Second, the visual pathway impact analysis shows that reducing visual feature dimensionality to 128 resulted in only 0.3 dB degradation, validating the key NAS insight regarding visual redundancy. Third, the real-time viability of our model at 0.19 RTF on mobile processors establishes practical feasibility for on-device deployment scenarios. Fourth, the superiority over random searchâ0.6 dB at equivalent parameter budgetsâunderscores the value of RL-guided architecture search. These comprehensive results establish that our NAS methodology successfully identifies high-performing architectures that random search would require thousands of iterations to discover. 5 Conclusion This paper presents a comprehensive reinforcement learning-based neural architecture search framework specifically tailored for audio-visual speech enhancement. By formulating architecture search as a constrained MDP with carefully designed state representations, action spaces, and multi-objective reward functions, our framework successfully bridges the efficiency gap between state-of-the-art performance and practical deployability on edge devices. The discovered 2.9-million-parameter model maintains 14.5 dB SI-SNR improvement, representing 95.4% of the baseline performance while achieving nine-fold compression, demonstrating efficiency-accuracy tradeoffs that align with emerging paradigms in resource-constrained machine learning [38, 52]. The research contributes several important findings to the multimodal neural architecture community. Our analysis demonstrates that visual lip-movement information can be represented in remarkably reduced dimensional spacesâjust 128 channels compared to 256 in baseline designsâwithout substantial quality loss, a finding that challenges conventional architectural assumptions about multimodal fusion [31]. The search process revealed that optimal TCN configurations employ three repeats of six-block structures with expansion ratio 2.0, suggesting that shallower but wider networks may represent a fundamental tradeoff in speech processing, consistent with findings in temporal sequence modeling [26]. Parameter redistribution patterns discovered through NAS suggest that concentrating learnable parameters in early processing stages enhances noise robustness compared to conventional designs. The additive fusion approach for multimodal combination outperformed concatenation-based alternatives while simultaneously reducing parameters, aligning with emerging attention-based multimodal fusion mechanisms [53, 54]. The resulting lightweight audio-visual speech enhancement modelâcomprising 11.7 MB of parametersâenables previously impractical deployment scenarios. Real-time hearing aids equipped with visual augmentation become feasible through the dramatic four-fold reduction in computational requirements. Mobile video conferencing applications can now incorporate audio-visual enhancement without imposing unacceptable battery drain. Always-on voice assistants with visual context awareness move from theoretical possibility to practical reality. These practical applications represent the ultimate motivation for this research: making powerful multimodal speech enhancement available to users on their personal devices, leveraging efficient inference techniques established in mobile computing [39, 40]. Future research should investigate several promising extensions of this framework. Dynamic architectures that adapt to changing acoustic conditions and noise characteristics could provide further efficiency gains through techniques like adaptive gating. Quantization-aware neural architecture search, combining this framework with low-bit representation learning [36] and pruning strategies [34, 35], could achieve additional compression beyond what is currently possible. Enhanced cross-modal attention mechanisms specifically designed for synchronizing audio-visual information [53, 55] could improve the quality of fusion while maintaining model compactness. Integration with self-supervised learning approaches could reduce dependence on carefully annotated training data and extend applicability to diverse real-world scenarios. Exploration of generative NAS approaches [21] and hardware-aware optimization strategies [44] will continue advancing the frontier of efficient multimodal speech processing. These extensions will build upon the principled multi-objective optimization framework we have established. The RL-NAS framework presented in this work establishes a generalizable blueprint for developing efficient multimodal systems that successfully balance high performance with stringent deployment constraints. By leveraging reinforcement learning to systematically explore constrained architectural spaces while optimizing multi-objective trade-offs, this approach demonstrates that heavyweight multimodal systems can be compressed without sacrificing fundamental quality. This methodology extends beyond speech enhancement to broader applications in multimodal processing, suggesting that hardware-aware NAS guided by reinforcement learning represents a powerful paradigm for realizing ubiquitous, intelligent voice interfaces. Acknowledgment This research acknowledges the support of UK Engineering and Physical Sciences Research Council (EPSRC) Grants Ref. EâP/Tâ021063/1EP/T021063/1 (COG-MHEAR). References [1] A. Gabbay, A. Ephrat, T. Halperin, and S. Peleg (2018) Seeing through noise: visually driven speaker separation and enhancement. External Links: 1708.06767, Link Cited by: §1, §2.4. [2] A. Gabbay, A. Shamir, and S. Peleg (2018) Visual speech enhancement. In Interspeech, Cited by: §1. [3] A. Ephrat et al. (2018) Looking to listen at the cocktail party: a visually-informed speaker separation network. arXiv:1804.02665. Cited by: §1. [4] Y. M. Assael, B. Shillingford, S. Whitehead, and N. de Freitas (2016) LipNet: end-to-end sentence-level lipreading. In arXiv:1611.01599, Cited by: §1. [5] J. S. Chung and A. Zisserman (2016) Lip reading sentences in the wild. arXiv:1611.05358. Cited by: §1. [6] X. Xu, Y. Wang, D. Xu, Y. Peng, C. Zhang, J. Jia, and B. Chen (2022) VSEGAN: visual speech enhancement generative adversarial network. External Links: 2102.02599, Link Cited by: §1, §2.3, §2.4. [7] A. Pandey and D. Wang (2019) TCNN: temporal convolutional neural network for real-time speech enhancement in the time domain. IEEE/ACM Transactions on Audio, Speech, and Language Processing 27, p. 6875â6879. External Links: Document Cited by: §1, §2.2, §2.6. [8] Y. Wang et al. (2018) Supervised speech separation based on deep learning: an overview. IEEE/ACM Transactions on Audio, Speech, and Language Processing 26 (10), p. 1795â1809. Cited by: §1. [9] Y. Wang et al. (2018) End-to-end speech separation by speaker-aware generative adversarial network. arXiv:1804.10848. Cited by: §1. [10] Y. C. Subakan et al. (2021) Attention is all you need in speech separation. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 51â55. Cited by: §1. [11] F. G. Germain et al. (2016) Speech enhancement using a two-stage network for student-teacher learning. arXiv:1605.06437. Cited by: §1. [12] H. Zhao et al. (2020) Monaural speech enhancement using dynamic layered frequency masking. In Interspeech, Cited by: §1. [13] B. Zoph and Q. V. Le (2017) Neural architecture search with reinforcement learning. External Links: 1611.01578, Link Cited by: §2.1. [14] E. Real, A. Aggarwal, Y. Huang, and Q. V. Le (2019) Regularized evolution for image classifier architecture search. arXiv:1802.01548. Cited by: §2.1. [15] H. Liu, K. Simonyan, and Y. Yang (2019) DARTS: differentiable architecture search. External Links: 1806.09055, Link Cited by: §2.1. [16] H. Pham, M. Guan, B. Zoph, Q. Le, and J. Dean (2018) Efficient neural architecture search via parameters sharing. In Proceedings of the 35th International Conference on Machine Learning, J. Dy and A. Krause (Eds.), Proceedings of Machine Learning Research, Vol. 80, p. 4095â4104. Cited by: §2.1. [17] H. Cai, L. Zhu, and S. Han (2019) ProxylessNAS: direct neural architecture search on target task and hardware. arXiv:1812.00332. Cited by: §2.1. [18] G. Wen, Z. Li, K. Azizzadenesheli, A. Anandkumar, and S. M. Benson (2021) Neural architecture search for predictor-based methods. Neurocomputing 424, p. 1â14. External Links: Link Cited by: §2.1. [19] T. Yamasaki et al. (2025) RBFleX-nas: training-free neural architecture search using radial basis function kernel. IEEE Transactions on Neural Networks and Learning Systems 36 (6). Cited by: §2.1. [20] T. Elsken, J. H. Metzen, and F. Hutter (2019) Efficient multi-objective neural architecture search via lamarckian evolution. International Conference on Learning Representations (ICLR). Cited by: §2.1, §3. [21] X. Zhai et al. (2025) Generative neural architecture search. Neurocomputing 642, p. 130360. Cited by: §2.1, §5. [22] X. Zhou, X. Wu, L. Feng, Z. Lu, and K. C. Tan (2025) Design principle transfer in neural architecture search via large language models. Proceedings of the AAAI Conference on Artificial Intelligence 39 (21), p. 23000â23008. Cited by: §2.1. [23] H. Zheng et al. (2021) Efficient neural architecture search for end-to-end speech recognition. In IEEE Spoken Language Technology Workshop (SLT), Cited by: §2.2. [24] S. Hu et al. (2022) Neural architecture search for time delay neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, p. 2180â2192. Cited by: §2.2. [25] N. C. Hoang, T. N. L. Nguyen, T. K. Doan, and Q. C. Nguyen (2024) Multi-stage temporal representation learning via global and local perspectives for real-time speech enhancement. Applied Acoustics 223, p. 110067. External Links: ISSN 0003-682X, Document, Link Cited by: §2.3. [26] S. Bai, J. Z. Kolter, and V. Koltun (2018) An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv:1803.01271. Cited by: §2.3, §5. [27] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager (2017) Temporal convolutional networks for action segmentation and detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2.3. [28] F. Yu and V. Koltun (2015) Multi-scale context aggregation by dilated convolutions. arXiv:1511.07122. Cited by: §2.3. [29] X. Wu et al. (2021) Audio-visual neural architecture search for noise-robust speech enhancement. IEEE Transactions on Multimedia 24, p. 312â325. Cited by: §2.3, §2.4. [30] T. Zhang et al. (2025) Multi-task neural architecture search using architecture embedding and transfer rank. arXiv:2504.00772. Cited by: §2.4. [31] I. Baltruschat et al. (2013) Multimodal learning with deep boltzmann machines. arXiv:1301.3568. Cited by: §2.4, §5. [32] J. Ngiam et al. (2011) Multimodal learning with deep boltzmann machines. arXiv:1301.3568. Cited by: §2.4. [33] E. Tsimperidis et al. (2020) Late fusion via subspace search with consistency preservation. In Interspeech, Cited by: §2.4. [34] S. Han, H. Mao, and W. J. Dally (2015) Deep compression: compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv:1510.00149. Cited by: §2.4, §5. [35] H. Zhou, J. M. Alvarez, and F. Porikli (2016) Learning to prune deep neural networks via layer-wise optimal brain surgeon. arXiv:1608.04228. Cited by: §2.4, §5. [36] B. Jacob et al. (2018) Quantization and training of neural networks for efficient integer-arithmetic-only inference. arXiv:1806.08342. Cited by: §2.4, §5. [37] Y. Lin et al. (2020) Neural network compression and acceleration. arXiv:2002.08679. Cited by: §2.4. [38] A. G. Howard et al. (2017) MobileNets: efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861. Cited by: §2.4, §5. [39] M. Sandler et al. (2018) MobileNetV2: inverted residuals and linear bottlenecks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2.4, §5. [40] Z. Chen et al. (2020) MobileBERT: a compact task-agnostic bert for resource-limited devices. In Annual Meeting of the Association for Computational Linguistics, Cited by: §2.4, §5. [41] S. Siddiqui et al. (2025) Efficient global neural architecture search via macro-micro space optimization. arXiv:2502.03553. Cited by: §2.4. [42] I. Goodfellow et al. (2014) Generative adversarial networks. arXiv:1406.2661. Cited by: §2.5. [43] P. Isola et al. (2017) Image-to-image translation with conditional adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2.5. [44] Y. Wang et al. (2024) Hardware-aware neural architecture search for edge devices. arXiv:2404.12456. Cited by: §2.6, §5. [45] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv:1707.06347. Cited by: §3.4. [46] V. Mnih et al. (2016) Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning, Cited by: §3.4. [47] T. Lillicrap et al. (2015) Continuous control with deep reinforcement learning. arXiv:1509.02971. Cited by: §3.4. [48] T. Afouras, J. S. Chung, and A. Zisserman (2018) LRS3-TED: a large-scale dataset for visual speech recognition. CoRR abs/1809.00496. External Links: Link, 1809.00496 Cited by: §4.1. [49] V. Panayotov et al. (2015) LibriSpeech: an asr corpus based on public domain audio books. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Cited by: §4.1. [50] M. A. Akeroyd, W. Bailey, J. Barker, T. J. Cox, J. F. Culling, S. Graetzer, G. Naylor, Z. PodwiĆska, and Z. Tu (2023) The 2nd clarity enhancement challenge for hearing aid speech intelligibility enhancement: overview and outcomes. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 1â5. External Links: Document Cited by: §4.1. [51] D. Snyder et al. (2018) Speaker recognition for the jsalt 2017 multilingual speaker recognition challenge. In Odyssey, Cited by: §4.1. [52] M. Tan and Q. V. Le (2019) EfficientNet: rethinking model scaling for convolutional neural networks. arXiv:1905.11946. Cited by: §5. [53] A. Vaswani et al. (2017) Attention is all you need. Advances in Neural Information Processing Systems. Cited by: §5, §5. [54] L. Lu et al. (2020) Transformer-based acoustic modeling for hybrid speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6124â6128. Cited by: §5. [55] X. Tan et al. (2019) Speech-transformer: a no-recurrence sequence-to-sequence model for speech recognition and translation. arXiv:1910.09799. Cited by: §5.