Nikhil Narasimhan

Srija Anand, Ashwin Sankar, Ishvinder Sethi et al.

Crowdsourced pairwise evaluation has emerged as a scalable approach for assessing foundation models. However, applying it to Text to Speech(TTS) introduces high variance due to linguistic diversity and multidimensional nature of speech perception. We present a controlled multidimensional pairwise evaluation framework for multilingual TTS that combines linguistic control with perceptually grounded annotation. Using 5K+ native and code-mixed sentences across 10 Indic languages, we evaluate 7 state-of-the-art TTS systems and collect over 120K pairwise comparisons from over 1900 native raters. In addition to overall preference, raters provide judgments across 6 perceptual dimensions: intelligibility, expressiveness, voice quality, liveliness, noise, and hallucinations. Using Bradley-Terry modeling, we construct a multilingual leaderboard, interpret human preference using SHAP analysis and analyze leaderboard reliability alongside model strengths and trade-offs across perceptual dimensions.

Ashwin Sankar, Sparsh Jain, Nikhil Narasimhan et al. · microsoft-research

Speech translation for Indian languages remains a challenging task due to the scarcity of large-scale, publicly available datasets that capture the linguistic diversity and domain coverage essential for real-world applications. Existing datasets cover a fraction of Indian languages and lack the breadth needed to train robust models that generalize beyond curated benchmarks. To bridge this gap, we introduce BhasaAnuvaad, the largest speech translation dataset for Indian languages, spanning over 44 thousand hours of audio and 17 million aligned text segments across 14 Indian languages and English. Our dataset is built through a threefold methodology: (a) aggregating high-quality existing sources, (b) large-scale web crawling to ensure linguistic and domain diversity, and (c) creating synthetic data to model real-world speech disfluencies. Leveraging BhasaAnuvaad, we train IndicSeamless, a state-of-the-art speech translation model for Indian languages that performs better than existing models. Our experiments demonstrate improvements in the translation quality, setting a new standard for Indian language speech translation. We will release all the code, data and model weights in the open-source, with permissive licenses to promote accessibility and collaboration.

Sambartha Ray Barman, Andrey Starenky, Sofia Bodnar et al.

Every major AI memory system in production today organises information by meaning. That organisation enables generalisation, analogy, and conceptual retrieval -- but it comes at a price. We prove that the same geometric structure enabling semantic generalisation makes interference, forgetting, and false recall inescapable. We formalise this tradeoff for \textit{semantically continuous kernel-threshold memories}: systems whose retrieval score is a monotone function of an inner product in a semantic feature space with finite local intrinsic dimension. Within this class we derive four results: (1) semantically useful representations have finite effective rank; (2) finite local dimension implies positive competitor mass in retrieval neighbourhoods; (3) under growing memory, retention decays to zero, yielding power-law forgetting curves under power-law arrival statistics; (4) for associative lures satisfying a $δ$-convexity condition, false recall cannot be eliminated by threshold tuning. We test these predictions across five architectures: vector retrieval, graph memory, attention-based context, BM25 filesystem retrieval, and parametric memory. Pure semantic systems express the vulnerability directly as forgetting and false recall. Reasoning-augmented systems partially override these symptoms but convert graceful degradation into catastrophic failure. Systems that escape interference entirely do so by sacrificing semantic generalisation. The price of meaning is interference, and no architecture we tested avoids paying it.

Sambartha Ray Barman, Andrey Starenky, Sophia Bodnar et al.

Why do we forget? Why do we remember things that never happened? The conventional answer points to biological hardware. We propose a different one: geometry. Here we show that high-dimensional embedding spaces, subjected to noise, interference, and temporal degradation, reproduce quantitative signatures of human memory with no phenomenon-specific engineering. Power-law forgetting ($b = 0.460 \pm 0.183$, human $b \approx 0.5$) arises from interference among competing memories, not from decay. The identical decay function without competitors yields $b \approx 0.009$, fifty times smaller. Time alone does not produce forgetting in this system. Competition does. Production embedding models (nominally 384--1{,}024 dimensions) concentrate their variance in only ${\sim}16$ effective dimensions, placing them deep in the interference-vulnerable regime. False memories require no engineering at all: cosine similarity on unmodified pre-trained embeddings reproduces the Deese--Roediger--McDermott false alarm rate ($0.583$ versus human ${\sim}0.55$) with zero parameter tuning and no boundary conditions. We did not build a false memory system. We found one already present in the raw geometry of semantic space. These results suggest that core memory phenomena are not bugs of biological implementation but features of any system that organizes information by meaning and retrieves it by proximity.