Smart glasses are increasingly positioned as the nextgeneration interface for ubiquitous access to large language models (LLMs). Nevertheless, achieving reliable interaction in real-world noisy environments remains a major challenge, particularly due to interference from side speech. In this work, we introduce a novel side-talk rejection multimicrophone Whisper (MMW) framework for smart glasses, incorporating three key innovations. First, we propose a Mix Block based on a Tri-Mamba architecture to effectively fuse multi-channel audio at the raw waveform level, while maintaining compatibility with streaming processing. Second, we design a Frame Diarization Mamba Layer to enhance frame-level side-talk suppression, facilitating more efficient fine-tuning of Whisper models. Third, we employ a Multi-Scale Group Relative Policy Optimization (GRPO) strategy to jointly optimize frame-level and utterance-level side speech suppression. Experimental evaluations demonstrate that the proposed MMW system can reduce the word error rate (WER) by 4.95% in noisy conditions.

Since Large Language Models or LLMs have demonstrated high-quality performance on many complex language tasks, there is a great interest in bringing these LLMs to mobile devices for faster responses and better privacy protection. However, the size of LLMs (i.e., billions of parameters) requires highly effective compression to fit into storage-limited devices. Among many compression techniques, weight-clustering, a form of non-linear quantization, is one of the leading candidates for LLM compression, and supported by modern smartphones. Yet, its training overhead is prohibitively significant for LLM fine-tuning. Especially, Differentiable KMeans Clustering, or DKM, has shown the state-of-the-art trade-off between compression ratio and accuracy regression, but its large memory complexity makes it nearly impossible to apply to train-time LLM compression. In this letter, we propose a memory-efficient DKM implementation, eDKM powered by novel techniques to reduce the memory footprint of DKM by orders of magnitudes. For a given tensor to be saved on CPU for the backward pass of DKM, we compressed the tensor by applying uniquification and sharding after checking if there is no duplicated tensor previously copied to CPU. Our experimental results demonstrate that eDKM can fine-tune and compress a pretrained LLaMA 7B model from 12.6 GB to $2.5 \mathrm{~GB}(3 \mathrm{~b} /$ weight) with the Alpaca dataset by reducing the train-time memory footprint of a decoder layer by $130 \times$, while delivering good accuracy on broader LLM benchmarks (i.e., $77.7 \%$ for PIQA, $66.1 \%$ for Winograde, and so on).

Device-directed speech detection (DDSD) is the binary classification task of distinguishing between queries directed at a voice assistant versus side conversation or background speech. State-of-the-art DDSD systems use verbal cues, e.g acoustic, text and/or automatic speech recognition system (ASR) features, to classify speech as device-directed or otherwise, and often have to contend with one or more of these modalities being unavailable when deployed in real-world settings. In this paper, we investigate fusion schemes for DDSD systems that can be made more robust to missing modalities. Concurrently, we study the use of non-verbal cues, specifically prosody features, in addition to verbal cues for DDSD. We present different approaches to combine scores and embeddings from prosody with the corresponding verbal cues, finding that prosody improves DDSD performance by upto 8.5% in terms of false acceptance rate (FA) at a given fixed operating point via non-linear intermediate fusion, while our use of modality dropout techniques improves the performance of these models by 7.4% in terms of FA when evaluated with missing modalities during inference time.

Streaming neural network models for fast frame-wise responses to various speech and sensory signals are widely adopted on resource-constrained platforms. Hence, increasing the learning capacity of such streaming models (i.e., by adding more parameters) to improve the predictive power may not be viable for real-world tasks. In this work, we propose a new loss, Streaming Anchor Loss (SAL), to better utilize the given learning capacity by encouraging the model to learn more from essential frames. More specifically, our SAL and its focal variations dynamically modulate the frame-wise cross entropy loss based on the importance of the corresponding frames so that a higher loss penalty is assigned for frames within the temporal proximity of semantically critical events. Therefore, our loss ensures that the model training focuses on predicting the relatively rare but task-relevant frames. Experimental results with standard lightweight convolutional and recurrent streaming networks on three different speech based detection tasks demonstrate that SAL enables the model to learn the overall task more effectively with improved accuracy and latency, without any additional data, model parameters, or architectural changes.

Suppressing unintended invocation of the device because of the speech that sounds like wake-word, or accidental button presses, is critical for a good user experience, and is referred to as False-Trigger-Mitigation (FTM). In case of multiple invocation options, the traditional approach to FTM is to use invocation-specific models, or a single model for all invocations. Both approaches are sub-optimal: the memory cost for the former approach grows linearly with the number of invocation options, which is prohibitive for on-device deployment, and does not take advantage of shared training data; while the latter is unable to accurately capture acoustic differences across different invocation types. To this end, we propose a Unified Acoustic Detector (UAD) for FTM when multiple invocation options are available on device. The proposed UAD is trained using a multi-task learning framework, where a jointly trained acoustic encoder model is augmented with invocation-specific classification layers. In the context of the FTM task, we show for the first time that using the shared model architecture across invocations (thus, keeping the model size similar to that of a monolithic model used for a single invocation type), we can not only match but largely improve the accuracy of the invocation-specific models. In particular, in the challenging case of touch-based invocation, we obtain 50% and 35% relative improvement in false positive rate at 99% true positive rate, when compared with a singleoutput model for both invocations, and separate models per invocation, respectively. Furthermore, we propose streaming and non-streaming variants of the UAD, and show that they both outperform a traditional ASR-based approach to FTM.

The adoption of multimodal interactions by Voice Assistants (VAs) is growing rapidly to enhance human-computer interactions.Smartwatches have now incorporated trigger-less methods of invoking VAs, such as Raise To Speak (RTS), where the user raises their watch and speaks to VAs without an explicit trigger.Current state-of-the-art RTS systems rely on heuristics and engineered Finite State Machines to fuse gesture and audio data for multimodal decision-making.However, these methods have limitations, including limited adaptability, scalability, and induced human biases.In this work, we propose a neural network based audio-gesture multimodal fusion system that (1) Better understands temporal correlation between audio and gesture data, leading to precise invocations (2) Generalizes to a wide range of environments and scenarios (3) Is lightweight and deployable on low-power devices, such as smartwatches, with quick launch times (4) Improves productivity in asset development processes.

DNN pruning is a popular way to reduce the size of a model, improve the inference latency, and minimize the power consumption on DNN accelerators. However, existing approaches might be too complex, expensive or ineffective to apply to a variety of vision/language tasks, DNN architectures and to honor structured pruning constraints. In this paper, we propose an efficient yet effective train-time pruning scheme, Parameter-free Differentiable Pruning (PDP), which offers state- of-the-art qualities in model size, accuracy, and training cost. PDP uses a dynamic function of weights during training to generate soft pruning masks for the weights in a parameter-free manner for a given pruning target. While differentiable, the simplicity and efficiency of PDP make it universal enough to deliver state-of-the-art random/structured/channel pruning results on various vision and natural language tasks. For example, for MobileNet-v1, PDP can achieve 68.2% top-1 ImageNet1k accuracy at 86.6% sparsity, which is 1.7% higher accuracy than those from the state-of-the-art algorithms. Also, PDP yields over 83.1% accuracy on Multi-Genre Natural Language Inference with 90% sparsity for BERT, while the next best from the existing techniques shows 81.5% accuracy. In addition, PDP can be applied to structured pruning, such as N:M pruning and channel pruning. For 1:4 structured pruning of ResNet18, PDP improved the top-1 ImageNet1k accuracy by over 3.6% over the state-of-the-art. For channel pruning of ResNet50, PDP reduced the top-1 ImageNet1k accuracy by 0.6% from the state-of-the-art.

We address the problem of detecting speech directed to a device that does not contain a specific wake-word.Specifically, we focus on audio coming from a touch-based invocation.Mitigating virtual assistants (VAs) activation due to accidental button presses is critical for user experience.While the majority of approaches to false trigger mitigation (FTM) are designed to detect the presence of a target keyword, inferring user intent in absence of keyword is difficult.This also poses a challenge when creating the training/evaluation data for such systems due to inherent ambiguity in the user's data.To this end, we propose a novel FTM approach that uses weakly-labeled training data obtained with a newly introduced data sampling strategy.While this sampling strategy reduces data annotation efforts, the data labels are noisy as the data are not annotated manually.We use these data to train an acoustics-only model for the FTM task by regularizing its loss function via knowledge distillation from an ASR-based (LatticeRNN) model.This improves the model decisions, resulting in 66% gain in accuracy, as measured by equal-error-rate (EER), over the base acoustics-only model.We also show that the ensemble of the LatticeRNN and acousticdistilled models brings further accuracy improvement of 20%.

Voice trigger detection is an important task, which enables activating a voice assistant when a target user speaks a keyword phrase.A detector is typically trained on speech data independent of speaker information and used for the voice trigger detection task.However, such a speaker independent voice trigger detector typically suffers from performance degradation on speech from underrepresented groups, such as accented speakers.In this work, we propose a novel voice trigger detector that can use a small number of utterances from a target speaker to improve detection accuracy.Our proposed model employs an encoder-decoder architecture.While the encoder performs speaker independent voice trigger detection, similar to the conventional detector, the decoder is trained with metric learning and predicts a personalized embedding for each utterance.A personalized voice trigger score is then obtained as a similarity score between the embeddings of enrollment utterances and a test utterance.The personalized embedding allows adapting to target speaker's speech when computing the voice trigger score, hence improving voice trigger detection accuracy.Experimental results show that the proposed approach achieves a 38% relative reduction in a false rejection rate (FRR) compared to a baseline speaker independent voice trigger model.

We present a unified and hardware efficient architecture for two stage voice trigger detection (VTD) and false trigger mitigation (FTM) tasks. Two stage VTD systems of voice assistants can get falsely activated to audio segments acoustically similar to the trigger phrase of interest. FTM systems cancel such activations by using post trigger audio context. Traditional FTM systems rely on automatic speech recognition lattices which are computationally expensive to obtain on device. We propose a streaming transformer (TF) encoder architecture, which progressively processes incoming audio chunks and maintains audio context to perform both VTD and FTM tasks using only acoustic features. The proposed joint model yields an average 18% relative reduction in false reject rate (FRR) for the VTD task at a given false alarm rate. Moreover, our model suppresses 95% of the false triggers with an additional one second of post-trigger audio. Finally, on-device measurements show 32% reduction in runtime memory and 56% reduction in inference time compared to non-streaming version of the model.

Voice-triggered smart assistants often rely on detection of a trigger-phrase before they start listening for the user request. Mitigation of false triggers is an important aspect of building a privacy-centric non-intrusive smart assistant. In this paper, we address the task of false trigger mitigation (FTM) using a novel approach based on analyzing automatic speech recognition (ASR) lattices using graph neural networks (GNN). The proposed approach uses the fact that decoding lattice of a falsely triggered audio exhibits uncertainties in terms of many alternative paths and unexpected words on the lattice arcs as compared to the lattice of a correctly triggered audio. A pure trigger-phrase detector model doesn't fully utilize the intent of the user speech whereas by using the complete decoding lattice of user audio, we can effectively mitigate speech not intended for the smart assistant. We deploy two variants of GNNs in this paper based on 1) graph convolution layers and 2) self-attention mechanism respectively. Our experiments demonstrate that GNNs are highly accurate in FTM task by mitigating ~87% of false triggers at 99% true positive rate (TPR). Furthermore, the proposed models are fast to train and efficient in parameter requirements.

We consider the design of two-pass voice trigger detection systems. We focus on the networks in the second pass that are used to re-score candidate segments obtained from the first-pass. Our baseline is an acoustic model(AM), with BiLSTM layers, trained by minimizing the CTC loss. We replace the BiLSTM layers with self-attention layers. Results on internal evaluation sets show that self-attention networks yield better accuracy while requiring fewer parameters. We add an auto-regressive decoder network on top of the self-attention layers and jointly minimize the CTC loss on the encoder and the cross-entropy loss on the decoder. This design yields further improvements over the baseline. We retrain all the models above in a multi-task learning(MTL) setting, where one branch of a shared network is trained as an AM, while the second branch classifies the whole sequence to be true-trigger or not. Results demonstrate that networks with self-attention layers yield $\sim$60% relative reduction in false reject rates for a given false-alarm rate, while requiring 10% fewer parameters. When trained in the MTL setup, self-attention networks yield further accuracy improvements. On-device measurements show that we observe 70% relative reduction in inference time. Additionally, the proposed network architectures are $\sim$5X faster to train.