arXiv subjects

Sriram Ganapathy

Publications and source records attributed to Sriram Ganapathy.

At least 19 recordsLinked to original sources

arXiv2026

BiG-SURE - Bipartite Graph for Semantic Uncertainty and Reliability Estimation of LLMs

Reliable uncertainty estimation is a crucial requirement for deploying large language models (LLMs) and vision-language models (VLMs) in safety-critical settings, especially when the model parameters are not accessible (black-box). We propose BiG-SURE, an uncertainty estimator based on cross-temperature semantic agreement. The method samples low-temperature responses as stable semantic anchors and high-temperature responses as probes under meaning-preserving input transformations. It then constructs an anchor-probe Bipartite Graph (BiG) using NLI-based entailment scores and defines confidence through the normalized squared spectral energy of this matrix, with uncertainty given by its complement. This bipartite graph-based Semantic Uncertainty and Reliability Estimation (SURE) score measures whether high-temperature probes remain semantically aligned with the model's stable low-temperature belief or not. We evaluate BiG-SURE on text QA, multilingual QA, and multimodal QA tasks across multiple model families. In these experiments, BiG-SURE improves average abstention AUROC over prior black-box uncertainty estimators, while remaining simple, unsupervised, and applicable to black-box model settings.

cs.CL↗

arXiv2026

ULTRAS -- Unified Learning of Transformer Representations for Audio and Speech Signals

Self-supervised learning (SSL) has driven impressive advances in speech processing by adopting time-domain prediction objectives, while audio representation learning frameworks operate on time-frequency spectrograms. Models optimized for one paradigm struggle to transfer to the other, highlighting the need for a joint framework. We propose Unified Learning of Transformer Representations for Audio and Speech (ULTRAS), where the masking and predictive modeling is performed over long patches of the data. The model, based on the transformer architecture, encodes spectral-patches of log-mel spectrogram features. The predictive modeling of masked segments is performed on spectral and temporal targets using a combined loss-function, forcing the representations to encode time and frequency traits. Experiments are performed on a variety of speech and audio tasks, where we illustrate that the ULTRAS framework achieves improved performance over other established baselines.

eess.AS↗

arXiv2026

Textless and Non-Parallel Speech-to-Speech Emotion Style Transfer

Given a pair of source and reference speech recordings, speech-to-speech (S2S) emotion style transfer involves the generation of an output speech that mimics the emotion characteristics of the reference while preserving the content and speaker attributes of the source. In this paper, we propose a speech-to-speech zero-shot emotion style transfer framework, termed S2S Zero-shot Emotion Style Transfer (S2S-ZEST), that enables the transfer of emotional attributes from the reference to the source while retaining the speaker identity and speech content. The S2S-ZEST framework consists of an analysis-synthesis pipeline in which the analysis module extracts semantic tokens, speaker representations, and emotion embeddings from speech. Using these representations, a pitch contour estimator and a duration predictor are learned. Further, a synthesis module is designed to generate speech based on the input representations and the derived factors. The analysis-synthesis pipeline is trained using an auto-encoding objective to enable efficient resynthesis during inference. For S2S emotion style transfer, the emotion embedding extracted from the reference speech along with the remaining representations from the source speech are used in the synthesis module to generate the style-transferred speech. In our experiments, we evaluate the converted speech on content and speaker preservation (with respect to the source) as well as on the effectiveness of the emotion style transfer (with respect to the reference). The proposed framework demonstrates improved emotion style transfer performance over prior methods in a textless and non-parallel setting. We also illustrate the application of the proposed work for data augmentation in emotion recognition tasks.

eess.AS↗

arXiv2026

Benchmarking Humans and Machines on Complex Multilingual Speech Understanding Tasks

Auditory attention and selective phase-locking are central to human speech understanding in complex acoustic scenes and cocktail party settings, yet these capabilities in multilingual subjects remain poorly understood. While machine understanding of natural speech has advanced in recent years, questions persist about comprehension of overlapped and mixed-channel speech. We propose a systematic paradigm for studying humans and machines in speech question-answering tasks in multilingual settings with clean and mixed-channel speech. For human listeners, selective attention to a target speaker was significantly better in their native language (L1) than in their second language (L2). For machine listening, speech-based large language models (LLMs) match or exceed human performance in clean, single-speaker conditions but often struggle to selectively attend in two-speaker settings. These results reveal a key divergence: humans rely on attentional cues that are more streamlined in their native language, whereas LLMs default to parallel information extraction which exceed human skills.

eess.AS↗

arXiv2026

Benchmarking Speech Systems for Frontline Health Conversations: The DISPLACE-M Challenge

The DIarization and Speech Processing for LAnguage understanding in Conversational Environments - Medical (DISPLACE-M) challenge introduces a conversational AI benchmark for understanding goal-oriented, real-world medical dialogues. The challenge addresses multi-speaker interactions between frontline health workers and care seekers, characterized by spontaneous, noisy and overlapping speech. As part of the challenge, medical conversational dataset comprising 40 hours of development and 15 hours of blind evaluation recordings was released. We provided baseline systems across 4 tasks - speaker diarization, automatic speech recognition, topic identification and dialogue summarization - to enable consistent benchmarking. System performance is evaluated using diarization error rate (DER), time-constrained minimum-permutation word error rate (tcpWER) and ROUGE-L. This paper describes the Phase-I evaluation - data, tasks and baseline systems - along with the summary of the evaluation results.

eess.AS↗

arXiv2026

A Mixture-of-Experts Model for Multimodal Emotion Recognition in Conversations

Emotion Recognition in Conversations (ERC) presents unique challenges, requiring models to capture the temporal flow of multi-turn dialogues and to effectively integrate cues from multiple modalities. We propose Mixture of Speech-Text Experts for Recognition of Emotions (MiSTER-E), a modular Mixture-of-Experts (MoE) framework designed to decouple two core challenges in ERC: modality-specific context modeling and multimodal information fusion. MiSTER-E leverages large language models (LLMs) fine-tuned for both speech and text to provide rich utterance-level embeddings, which are then enhanced through a convolutional-recurrent context modeling layer. The system integrates predictions from three experts-speech-only, text-only, and cross-modal-using a learned gating mechanism that dynamically weighs their outputs. To further encourage consistency and alignment across modalities, we introduce a supervised contrastive loss between paired speech-text representations and a KL-divergence-based regulariza-tion across expert predictions. Importantly, MiSTER-E does not rely on speaker identity at any stage. Experiments on three benchmark datasets-IEMOCAP, MELD, and MOSI-show that our proposal achieves 70.9%, 69.5%, and 87.9% weighted F1-scores respectively, outperforming several baseline speech-text ERC systems. We also provide various ablations to highlight the contributions made in the proposed approach.

cs.CL↗

arXiv2026

FESTA: Functionally Equivalent Sampling for Trust Assessment of Multimodal LLMs

The accurate trust assessment of multimodal large language models (MLLMs) generated predictions, which can enable selective prediction and improve user confidence, is challenging due to the diverse multi-modal input paradigms. We propose Functionally Equivalent Sampling for Trust Assessment (FESTA), a multimodal input sampling technique for MLLMs, that generates an uncertainty measure based on the equivalent and complementary input samplings. The proposed task-preserving sampling approach for uncertainty quantification expands the input space to probe the consistency (through equivalent samples) and sensitivity (through complementary samples) of the model. FESTA uses only input-output access of the model (black-box), and does not require ground truth (unsupervised). The experiments are conducted with various off-the-shelf multi-modal LLMs, on both visual and audio reasoning tasks. The proposed FESTA uncertainty estimate achieves significant improvement (33.3% relative improvement for vision-LLMs and 29.6% relative improvement for audio-LLMs) in selective prediction performance, based on area-under-receiver-operating-characteristic curve (AUROC) metric in detecting mispredictions. The code implementation is open-sourced.

cs.AI↗

arXiv2025

Leveraging Content and Acoustic Representations for Speech Emotion Recognition

Speech emotion recognition (SER), the task of identifying the expression of emotion from spoken content, is challenging due to the difficulty in extracting representations that capture emotional attributes from speech. The scarcity of labeled datasets further complicates the challenge where large models are prone to over-fitting. In this paper, we propose CARE (Content and Acoustic Representations of Emotions), where we design a dual encoding scheme which emphasizes semantic and acoustic factors of speech. While the semantic encoder is trained using distillation from utterance-level text representations, the acoustic encoder is trained to predict low-level frame-wise features of the speech signal. The proposed dual encoding scheme is a base-sized model trained only on unsupervised raw speech. With a simple light-weight classification model trained on the downstream task, we show that the CARE embeddings provide effective emotion recognition on a variety of datasets. We compare the proposal with several other self-supervised models as well as recent large-language model based approaches. In these evaluations, the proposed CARE is shown to be the best performing model based on average performance across 8 diverse datasets. We also conduct several ablation studies to analyze the importance of various design choices.

eess.AS↗

arXiv2025

Uncovering the role of semantic and acoustic cues in normal and dichotic listening

Speech comprehension is an involuntary task for the healthy human brain, yet the understanding of the mechanisms underlying this brain functionality remains obscure. In this paper, we aim to quantify the role of acoustic and semantic information streams in complex listening conditions. We propose a paradigm to understand the encoding of the speech cues in electroencephalogram (EEG) data, by designing a match-mismatch (MM) classification task. The MM task involves identifying whether the stimulus (speech) and response (EEG) correspond to each other. We build a multimodal deep-learning based sequence model STEM, which is input with acoustic stimulus (speech envelope), semantic stimulus (textual representations of speech), and the neural response (EEG data). We perform extensive experiments on two separate conditions, i) natural passive listening and, ii) a dichotic listening requiring auditory attention. Using the MM task as the analysis framework, we observe that - a) speech perception is fragmented based on word boundaries, b) acoustic and semantic cues offer similar levels of MM task performance in natural listening conditions, and c) semantic cues offer significantly improved MM classification over acoustic cues in dichotic listening task. The comparison of the STEM with previously proposed MM models shows significant performance improvements for the proposed approach. The analysis and understanding from this study allows the quantification of the roles played by acoustic and semantic cues in diverse listening tasks and in providing further evidences of right-ear advantage in dichotic listening.

eess.AS↗

arXiv2025

Towards the Next Frontier in Speech Representation Learning Using Disentanglement

The popular frameworks for self-supervised learning of speech representations have largely focused on frame-level masked prediction of speech regions. While this has shown promising downstream task performance for speech recognition and related tasks, this has largely ignored factors of speech that are encoded at coarser level, like characteristics of the speaker or channel that remain consistent through-out a speech utterance. In this work, we propose a framework for Learning Disentangled Self Supervised (termed as Learn2Diss) representations of speech, which consists of frame-level and an utterance-level encoder modules. The two encoders are initially learned independently, where the frame-level model is largely inspired by existing self supervision techniques, thereby learning pseudo-phonemic representations, while the utterance-level encoder is inspired by constrastive learning of pooled embeddings, thereby learning pseudo-speaker representations. The joint learning of these two modules consists of disentangling the two encoders using a mutual information based criterion. With several downstream evaluation experiments, we show that the proposed Learn2Diss achieves state-of-the-art results on a variety of tasks, with the frame-level encoder representations improving semantic tasks, while the utterance-level representations improve non-semantic tasks.

cs.CL↗

arXiv2025

ABHINAYA -- A System for Speech Emotion Recognition In Naturalistic Conditions Challenge

Speech emotion recognition (SER) in naturalistic settings remains a challenge due to the intrinsic variability, diverse recording conditions, and class imbalance. As participants in the Interspeech Naturalistic SER Challenge which focused on these complexities, we present Abhinaya, a system integrating speech-based, text-based, and speech-text models. Our approach fine-tunes self-supervised and speech large language models (SLLM) for speech representations, leverages large language models (LLM) for textual context, and employs speech-text modeling with an SLLM to capture nuanced emotional cues. To combat class imbalance, we apply tailored loss functions and generate categorical decisions through majority voting. Despite one model not being fully trained, the Abhinaya system ranked 4th among 166 submissions. Upon completion of training, it achieved state-of-the-art performance among published results, demonstrating the effectiveness of our approach for SER in real-world conditions.

cs.SD↗

arXiv2025

Benchmarking and Confidence Evaluation of LALMs For Temporal Reasoning

The popular success of text-based large language models (LLM) has streamlined the attention of the multimodal community to combine other modalities like vision and audio along with text to achieve similar multimodal capabilities. In this quest, large audio language models (LALMs) have to be evaluated on reasoning related tasks which are different from traditional classification or generation tasks. Towards this goal, we propose a novel dataset called temporal reasoning evaluation of audio (TREA). We benchmark open-source LALMs and observe that they are consistently behind human capabilities on the tasks in the TREA dataset. While evaluating LALMs, we also propose an uncertainty metric, which computes the invariance of the model to semantically identical perturbations of the input. Our analysis shows that the accuracy and uncertainty metrics are not necessarily correlated and thus, points to a need for wholesome evaluation of LALMs for high-stakes applications.

cs.CL↗

arXiv2025

Spoken Language Understanding on Unseen Tasks With In-Context Learning

Spoken language understanding (SLU) tasks involve diverse skills that probe the information extraction, classification and/or generation capabilities of models. In this setting, task-specific training data may not always be available. While traditional task-specific SLU models are unable to cater to such requirements, the speech-text large language models (LLMs) offer a promising alternative with emergent abilities. However, out of-the-box, our evaluations indicate that the zero/few-shot performance of prominent open-source speech-text LLMs on SLU tasks are not up to the mark. In this paper, we introduce a novel approach to robust task-agnostic fine-tuning using randomized class labels. With this proposed fine-tuning, we illustrate that the performance of the speech-text LLMs on an unseen task is significantly improved over standard approaches. Critically, the proposed approach avoids the requirement of task-specific data annotations for enabling new tasks in speech-text LLMs.

cs.CL↗

arXiv2025

LLM supervised Pre-training for Multimodal Emotion Recognition in Conversations

Emotion recognition in conversations (ERC) is challenging due to the multimodal nature of the emotion expression. In this paper, we propose to pretrain a text-based recognition model from unsupervised speech transcripts with LLM guidance. These transcriptions are obtained from a raw speech dataset with a pre-trained ASR system. A text LLM model is queried to provide pseudo-labels for these transcripts, and these pseudo-labeled transcripts are subsequently used for learning an utterance level text-based emotion recognition model. We use the utterance level text embeddings for emotion recognition in conversations along with speech embeddings obtained from a recently proposed pre-trained model. A hierarchical way of training the speech-text model is proposed, keeping in mind the conversational nature of the dataset. We perform experiments on three established datasets, namely, IEMOCAP, MELD, and CMU- MOSI, where we illustrate that the proposed model improves over other benchmarks and achieves state-of-the-art results on two out of these three datasets.

eess.AS↗

arXiv2024

End-to-End Supervised Hierarchical Graph Clustering for Speaker Diarization

Speaker diarization, the task of segmenting an audio recording based on speaker identity, constitutes an important speech pre-processing step for several downstream applications.The conventional approach to diarization involves multiple steps of embedding extraction and clustering, which are often optimized in an isolated fashion. While end-to-end diarization systems attempt to learn a single model for the task, they are often cumbersome to train and require large supervised datasets. In this paper, we propose an end-to-end supervised hierarchical clustering algorithm based on graph neural networks (GNN), called End-to-end Supervised HierARchical Clustering (E-SHARC). The embedding extractor is initialized using a pre-trained x-vector model while the GNN model is trained initially using the x-vector embeddings from the pre-trained model. Finally, the E-SHARC model uses the front-end mel-filterbank features as input and jointly optimizes the embedding extractor and the GNN clustering module, performing representation learning, metric learning, and clustering with end-to-end optimization. Further, with additional inputs from an external overlap detector, the E-SHARC approach is capable of predicting the speakers in the overlapping speech regions. The experimental evaluation on benchmark datasets like AMI, Voxconverse and DISPLACE, illustrates that the proposed E-SHARC framework provides competitive diarization results using graph based clustering methods.

eess.AS↗

arXiv2024

Gradient-free Post-hoc Explainability Using Distillation Aided Learnable Approach

The recent advancements in artificial intelligence (AI), with the release of several large models having only query access, make a strong case for explainability of deep models in a post-hoc gradient free manner. In this paper, we propose a framework, named distillation aided explainability (DAX), that attempts to generate a saliency-based explanation in a model agnostic gradient free application. The DAX approach poses the problem of explanation in a learnable setting with a mask generation network and a distillation network. The mask generation network learns to generate the multiplier mask that finds the salient regions of the input, while the student distillation network aims to approximate the local behavior of the black-box model. We propose a joint optimization of the two networks in the DAX framework using the locally perturbed input samples, with the targets derived from input-output access to the black-box model. We extensively evaluate DAX across different modalities (image and audio), in a classification setting, using a diverse set of evaluations (intersection over union with ground truth, deletion based and subjective human evaluation based measures) and benchmark it with respect to $9$ different methods. In these evaluations, the DAX significantly outperforms the existing approaches on all modalities and evaluation metrics.

cs.AI↗

arXiv2024

STAB: Speech Tokenizer Assessment Benchmark

Representing speech as discrete tokens provides a framework for transforming speech into a format that closely resembles text, thus enabling the use of speech as an input to the widely successful large language models (LLMs). Currently, while several speech tokenizers have been proposed, there is ambiguity regarding the properties that are desired from a tokenizer for specific downstream tasks and its overall generalizability. Evaluating the performance of tokenizers across different downstream tasks is a computationally intensive effort that poses challenges for scalability. To circumvent this requirement, we present STAB (Speech Tokenizer Assessment Benchmark), a systematic evaluation framework designed to assess speech tokenizers comprehensively and shed light on their inherent characteristics. This framework provides a deeper understanding of the underlying mechanisms of speech tokenization, thereby offering a valuable resource for expediting the advancement of future tokenizer models and enabling comparative analysis using a standardized benchmark. We evaluate the STAB metrics and correlate this with downstream task performance across a range of speech tasks and tokenizer choices.

cs.CL↗

arXiv2024

Improving Self-supervised Pre-training using Accent-Specific Codebooks

Speech accents present a serious challenge to the performance of state-of-the-art end-to-end Automatic Speech Recognition (ASR) systems. Even with self-supervised learning and pre-training of ASR models, accent invariance is seldom achieved. In this work, we propose an accent-aware adaptation technique for self-supervised learning that introduces a trainable set of accent-specific codebooks to the self-supervised architecture. These learnable codebooks enable the model to capture accent specific information during pre-training, that is further refined during ASR finetuning. On the Mozilla Common Voice dataset, our proposed approach outperforms all other accent-adaptation approaches on both seen and unseen English accents, with up to 9% relative reduction in word error rate (WER).

cs.CL↗