The AI @ ADI Reading Digest: Recommended paper and research about AI at the edge.

📚 The AI @ ADI Reading Digest: Recommended paper and research about AI at the edge.

At ADI, AI means developing world-class models that overcome latency, power, and cost constraints to deliver insights. It’s the end of January, and we are all kicking of our plans and laying down our roadmaps. With winter around us, it is the perfect time for hot cocoa, a warm fireplace, and digging into the latest thought-provoking AI research recommended by our team:

📄 OpenTSLM: Time Series Language Models for Reasoning over Multivariate Medical Time and Text data

👍🏼 Recommended by: Sambarta Dasgupta

💡 Why it’s worth reading:  The paper proposes a multi-modal time series reasoning model, with the two data modalities being time series and text. The paper focuses on the healthcare domain - where the time series can be various vital stats; the text is the user prompt, providing additional context and asking questions. The authors present two model architectures - one based on concatenating time and text embeddings, with the other one based on cross-attention. The model uses LLM backbones like Llama and Gemma - and low rank approximation for fine-tuning. They authors also used LLMs to annotate the data, with a subset reviewed by human experts. Their resulting model outperforms some of the SOTA over the benchmark datasets. Nonetheless, the authors concluded by identifying some of the gaps in their approach - patch encoding, zero shot evaluation on unknown dataset and so on. Besides this, one other area of improvement could have been adding a benchmark based on a combination of traditional statistical time series feature extractor and LLM reasoner.

📄 SincQDR-VAD: A Noise-Robust Voice Activity Detection Framework Leveraging Learnable Filters and Ranking-Aware Optimization

👍🏼 Recommended by: Shanza Iftikar

💡 Why it’s worth reading: This paper proposes a voice activity detection system designed to work reliably in very noisy environments. Instead of relying on fixed audio features like MFCCs, the model learns its own band-pass filters directly from raw audio using a SincNet-style front end. This lets the network adapt its frequency analysis to the data, making it more robust to noise and different acoustic conditions. Also, the paper introduces a ranking-based loss function called Quadratic Disparity Ranking (QDR). Unlike standard classification losses e.g. Binary cross entropy loss that treat each frame independently, QDR focuses on the relative ordering of predictions. It encourages the model to give consistently higher scores to speech frames than to non-speech frames, enforcing a margin between them. This is especially helpful for VAD, where non-speech frames heavily outnumber speech frames. By combining learnable filters (which improve feature robustness) with ranking-aware optimization (which improves decision consistency), SincQDR-VAD achieves stronger performance than traditional VAD methods, particularly in low-SNR and real-world noisy scenarios.