arXiv subjects

Santosh Kesiraju

Publications and source records attributed to Santosh Kesiraju.

At least 19 recordsLinked to original sources

arXiv2026

VākQA: A Benchmark and Evaluation Study for Telugu Spoken Factoid Question Answering

Question answering has advanced rapidly with large language models, but predominantly for high-resource languages, in both text and spoken settings. Spoken question answering (SQA) benchmark for Telugu remains unexplored, and the reliability of automatic evaluation in this setting remains unquantified. We introduce VākQA, a Telugu SQA benchmark of 2,001 factoid question-answer pairs across six domains, with 2.53 hours of speech audio, bilingual transcriptions, and human-verified reference answers. We first validate evaluation methods against human judgements: Gemini-as-a-judge best approximates human ratings but is non-uniformly strict, while open-weight judges systematically penalize correct Telugu answers that differ in surface form from the reference. Using this validated setup, we benchmark proprietary and open-weight models across input modality, language, and domain. We observe that Telugu phrasing retains cultural specificity that is lost in translation, speech input introduces phonetic confusions that alter question meaning, and cascaded ASR-MT errors compound progressively. VākQA is publicly released.

cs.CL↗

arXiv2026

S-DiverSe: Spanish Diverse Speech

Automatic speech recognition (ASR) has advanced remarkably for standard speech, yet speech affected by neurological conditions remains a challenge. We present S-DiverSe (Spanish Diverse Speech), a corpus of 3.2 hours of in-the-wild Spanish speech from 22 speakers with amyotrophic lateral sclerosis, Parkinson's disease, and stroke. The dataset contains 444 manually transcribed audio segments with metadata on speaker sex, disease type, and intelligibility. S-DiverSe is designed to support ASR evaluation and development for neurologically affected Spanish speech. We describe the dataset, analyze its composition, and report baseline ASR results alongside initial adaptation experiments. Our findings reveal that heuristic text post-processing is more robust than fine-tuning for out-of-domain neurological Spanish speech. This underscores the need for dedicated in-the-wild Spanish benchmarks.

cs.CL↗

arXiv2026

ORCA: Open-ended Response Correctness Assessment for Audio Question Answering

Reliable assessment of the abilities of large audio language models (LALMs) is essential to advancing the state of the art. As benchmarks rapidly evolve to incorporate complex reasoning and subjective tasks, they increasingly necessitate open-ended responses from LALMs. We present Open-ended Response Correctness Assessment (ORCA) -- a reliable and lightweight model-based approach for answer correctness and disagreement modeling. We employ a three-stage annotation pipeline combining human judgment, structured feedback, and human-AI correction, yielding 9,663 annotations across 3,699 question-answer pairs from 15 LALMs on three audio understanding and reasoning benchmarks (achieving a Krippendorff's alpha of 0.82). Our experiments employing curriculum learning show that ORCA models achieve a Spearman correlation of 0.91 with average human correctness ratings on seen benchmarks and generalize to unseen benchmarks with a score of 0.85, outperforming several LLM judge baselines including Gemini 2.5 Flash. Furthermore, we demonstrate that ORCA's predicted variance correlates strongly with human disagreement, allowing it to effectively identify problematic benchmark items.

cs.SD↗

arXiv2026

Robustness assessment of large audio language models in multiple-choice evaluation

Recent advances in large audio language models (LALMs) have primarily been assessed using a multiple-choice question answering (MCQA) framework. However, subtle changes, such as shifting the order of choices, result in substantially different results. Existing MCQA frameworks do not account for this variability and report a single accuracy number per benchmark or category. We dive into the MCQA evaluation framework and conduct a systematic study spanning three benchmarks (MMAU, MMAR and MMSU) and four models: Audio Flamingo 2, Audio Flamingo 3, Qwen2.5-Omni-7B-Instruct, and Kimi-Audio-7B-Instruct. Our findings indicate that models are sensitive not only to the ordering of choices, but also to the paraphrasing of the question and the choices. Finally, we propose a simpler evaluation protocol and metric that account for subtle variations and provide a more detailed evaluation report of LALMs within the MCQA framework.

cs.CL↗

arXiv2026

Joint Speech and Text Training for LLM-Based End-to-End Spoken Dialogue State Tracking

End-to-end spoken dialogue state tracking (DST) is made difficult by the tandem of having to handle speech input and data scarcity. Combining speech foundation encoders and large language models has been proposed in recent work as to alleviate some of this difficulty. Although this approach has been shown to result in strong spoken DST models, achieving state-of-the-art performance in realistic multi-turn DST, it struggles to generalize across domains and requires annotated spoken DST training data for each domain of interest. However, collecting such data for every target domain is both costly and difficult. Noting that textual DST data is more easily obtained for various domains, in this work, we propose jointly training on available spoken DST data and written textual data from other domains as a way to achieve cross-domain generalization. We conduct experiments which show the efficacy of our proposed method for getting good cross-domain DST performance without relying on spoken training data from the target domains.

cs.CL↗

arXiv2026

FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings

This paper presents factorized linear projection (FLiP) models for understanding pretrained sentence embedding spaces. We train FLiP models to recover the lexical content from multilingual (LaBSE), multimodal (SONAR) and API-based (Gemini) sentence embedding spaces in several high- and mid-resource languages. We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines. Using this as a diagnostic tool, we uncover the modality and language biases across the selected sentence encoders and provide practitioners with intrinsic insights about the encoders without relying on conventional downstream evaluation tasks. Our implementation is public https://github.com/BUTSpeechFIT/FLiP.

cs.CL↗

arXiv2026

FiLM-Based Speaker Conditioning of a SpeechLLM for Pathological Speech Recognition

Automatic speech recognition (ASR) has advanced remarkably for standard speech; however, pathological speech from neurological conditions remains a significant challenge. We investigate speaker conditioning via Feature-wise Linear Modulation (FiLM), injecting x-vector-derived information into each transformer layer of a frozen ASR encoder to adapt internal representations to individual pathological speakers without modifying base model weights. We benchmark this for the ASR task against standard and parameter-efficient fine-tuning baselines, complemented by post-processing, on Spanish and English pathological speech. Additionally, we evaluate if the adapted model preserves the ability to answer speech-related questions. Results show that speaker-conditioned ASR is competitive with established adaptation strategies while retaining performance on non-conditioned speech.

cs.CL↗

arXiv2025

MMAU-Pro: A Challenging and Comprehensive Benchmark for Holistic Evaluation of Audio General Intelligence

Audio comprehension-including speech, non-speech sounds, and music-is essential for achieving human-level intelligence. Consequently, AI agents must demonstrate holistic audio understanding to qualify as generally intelligent. However, evaluating auditory intelligence comprehensively remains challenging. To address this gap, we introduce MMAU-Pro, the most comprehensive and rigorously curated benchmark for assessing audio intelligence in AI systems. MMAU-Pro contains 5,305 instances, where each instance has one or more audios paired with human expert-generated question-answer pairs, spanning speech, sound, music, and their combinations. Unlike existing benchmarks, MMAU-Pro evaluates auditory intelligence across 49 unique skills and multiple complex dimensions, including long-form audio comprehension, spatial audio reasoning, multi-audio understanding, among others. All questions are meticulously designed to require deliberate multi-hop reasoning, including both multiple-choice and open-ended response formats. Importantly, audio data is sourced directly ``from the wild" rather than from existing datasets with known distributions. We evaluate 22 leading open-source and proprietary multimodal AI models, revealing significant limitations: even state-of-the-art models such as Gemini 2.5 Flash and Audio Flamingo 3 achieve only 59.2% and 51.7% accuracy, respectively, approaching random performance in multiple categories. Our extensive analysis highlights specific shortcomings and provides novel insights, offering actionable perspectives for the community to enhance future AI systems' progression toward audio general intelligence. The benchmark and code is available at https://sonalkum.github.io/mmau-pro.

eess.AS↗

arXiv2025

DeCRED: Decoder-Centric Regularization for Encoder-Decoder Based Speech Recognition

This paper presents a simple yet effective regularization for the internal language model induced by the decoder in encoder-decoder ASR models, thereby improving robustness and generalization in both in- and out-of-domain settings. The proposed method, Decoder-Centric Regularization in Encoder-Decoder (DeCRED), adds auxiliary classifiers to the decoder, enabling next token prediction via intermediate logits. Empirically, DeCRED reduces the mean internal LM BPE perplexity by 36.6% relative to 11 test sets. Furthermore, this translates into actual WER improvements over the baseline in 5 of 7 in-domain and 3 of 4 out-of-domain test sets, reducing macro WER from 6.4% to 6.3% and 18.2% to 16.2%, respectively. On TEDLIUM3, DeCRED achieves 7.0% WER, surpassing the baseline and encoder-centric InterCTC regularization by 0.6% and 0.5%, respectively. Finally, we compare DeCRED with OWSM v3.1 and Whisper-medium, showing competitive WERs despite training on much less data with fewer parameters.

eess.AS↗

arXiv2025

End-to-End Speech Translation for Low-Resource Languages Using Weakly Labeled Data

The scarcity of high-quality annotated data presents a significant challenge in developing effective end-to-end speech-to-text translation (ST) systems, particularly for low-resource languages. This paper explores the hypothesis that weakly labeled data can be used to build ST models for low-resource language pairs. We constructed speech-to-text translation datasets with the help of bitext mining using state-of-the-art sentence encoders. We mined the multilingual Shrutilipi corpus to build Shrutilipi-anuvaad, a dataset comprising ST data for language pairs Bengali-Hindi, Malayalam-Hindi, Odia-Hindi, and Telugu-Hindi. We created multiple versions of training data with varying degrees of quality and quantity to investigate the effect of quality versus quantity of weakly labeled data on ST model performance. Results demonstrate that ST systems can be built using weakly labeled data, with performance comparable to massive multi-modal multilingual baselines such as SONAR and SeamlessM4T.

cs.CL↗

arXiv2025

Approaching Dialogue State Tracking via Aligning Speech Encoders and LLMs

In this work, we approach spoken Dialogue State Tracking (DST) by bridging the representation spaces of speech encoders and LLMs via a small connector module, with a focus on fully open-sourced and open-data components (WavLM-large, OLMo). We focus on ablating different aspects of such systems including full/LoRA adapter fine-tuning, the effect of agent turns in the dialogue history, as well as fuzzy matching-based output post-processing, which greatly improves performance of our systems on named entities in the dialogue slot values. We conduct our experiments on the SpokenWOZ dataset, and additionally utilize the Speech-Aware MultiWOZ dataset to augment our training data. Ultimately, our best-performing WavLM + connector + OLMo-1B aligned models achieve state of the art on the SpokenWOZ test set (34.66% JGA), and our system with Gemma-2-9B-instruct further surpasses this result, reaching 42.17% JGA on SpokenWOZ test.

eess.AS↗

arXiv2025

Factors affecting the in-context learning abilities of LLMs for dialogue state tracking

This study explores the application of in-context learning (ICL) to the dialogue state tracking (DST) problem and investigates the factors that influence its effectiveness. We use a sentence embedding based k-nearest neighbour method to retrieve the suitable demonstrations for ICL. The selected demonstrations, along with the test samples, are structured within a template as input to the LLM. We then conduct a systematic study to analyse the impact of factors related to demonstration selection and prompt context on DST performance. This work is conducted using the MultiWoZ2.4 dataset and focuses primarily on the OLMo-7B-instruct, Mistral-7B-Instruct-v0.3, and Llama3.2-3B-Instruct models. Our findings provide several useful insights on in-context learning abilities of LLMs for dialogue state tracking.

cs.CL↗

arXiv2025

IIITH-BUT system for IWSLT 2025 low-resource Bhojpuri to Hindi speech translation

This paper presents the submission of IIITH-BUT to the IWSLT 2025 shared task on speech translation for the low-resource Bhojpuri-Hindi language pair. We explored the impact of hyperparameter optimisation and data augmentation techniques on the performance of the SeamlessM4T model fine-tuned for this specific task. We systematically investigated a range of hyperparameters including learning rate schedules, number of update steps, warm-up steps, label smoothing, and batch sizes; and report their effect on translation quality. To address data scarcity, we applied speed perturbation and SpecAugment and studied their effect on translation quality. We also examined the use of cross-lingual signal through joint training with Marathi and Bhojpuri speech data. Our experiments reveal that careful selection of hyperparameters and the application of simple yet effective augmentation techniques significantly improve performance in low-resource settings. We also analysed the translation hypotheses to understand various kinds of errors that impacted the translation quality in terms of BLEU.

cs.CL↗

arXiv2024

Aligning Pre-trained Models for Spoken Language Translation

This paper investigates a novel approach to end-to-end speech translation (ST) based on aligning frozen pre-trained automatic speech recognition (ASR) and machine translation (MT) models via a small connector module (Q-Former, our Subsampler-Transformer Encoder). This connector bridges the gap between the speech and text modalities, transforming ASR encoder embeddings into the latent representation space of the MT encoder while being the only part of the system optimized during training. Experiments are conducted on the How2 English-Portuguese dataset as we investigate the alignment approach in a small-scale scenario focusing on ST. While keeping the size of the connector module constant and small in comparison ( < 5% of the size of the larger aligned models), increasing the size and capability of the foundation ASR and MT models universally improves translation results. We also find that the connectors can serve as domain adapters for the foundation MT models, significantly improving translation performance in the aligned ST setting. We conclude that this approach represents a viable and scalable approach to training end-to-end ST systems.

cs.CL↗

arXiv2024

Improving Automatic Speech Recognition with Decoder-Centric Regularisation in Encoder-Decoder Models

This paper proposes a simple yet effective way of regularising the encoder-decoder-based automatic speech recognition (ASR) models that enhance the robustness of the model and improve the generalisation to out-of-domain scenarios. The proposed approach is dubbed as $\textbf{De}$coder-$\textbf{C}$entric $\textbf{R}$egularisation in $\textbf{E}$ncoder-$\textbf{D}$ecoder (DeCRED) architecture for ASR, where auxiliary classifier(s) is introduced in layers of the decoder module. Leveraging these classifiers, we propose two decoding strategies that re-estimate the next token probabilities. Using the recent E-branchformer architecture, we build strong ASR systems that obtained competitive WERs as compared to Whisper-medium and outperformed OWSM v3; while relying only on a fraction of training data and model size. On top of such a strong baseline, we show that DeCRED can further improve the results and, moreover, generalise much better to out-of-domain scenarios, where we show an absolute reduction of 2.7 and 2.9 WERs on AMI and Gigaspeech datasets, respectively. We provide extensive analysis and accompanying experiments that support the benefits of the proposed regularisation scheme.

eess.AS↗

arXiv2024

Beyond the Labels: Unveiling Text-Dependency in Paralinguistic Speech Recognition Datasets

Paralinguistic traits like cognitive load and emotion are increasingly recognized as pivotal areas in speech recognition research, often examined through specialized datasets like CLSE and IEMOCAP. However, the integrity of these datasets is seldom scrutinized for text-dependency. This paper critically evaluates the prevalent assumption that machine learning models trained on such datasets genuinely learn to identify paralinguistic traits, rather than merely capturing lexical features. By examining the lexical overlap in these datasets and testing the performance of machine learning models, we expose significant text-dependency in trait-labeling. Our results suggest that some machine learning models, especially large pre-trained models like HuBERT, might inadvertently focus on lexical characteristics rather than the intended paralinguistic features. The study serves as a call to action for the research community to reevaluate the reliability of existing datasets and methodologies, ensuring that machine learning models genuinely learn what they are designed to recognize.

eess.AS↗

arXiv2024

A Bayesian Multilingual Document Model for Zero-shot Topic Identification and Discovery

In this paper, we present a Bayesian multilingual document model for learning language-independent document embeddings. The model is an extension of BaySMM [Kesiraju et al 2020] to the multilingual scenario. It learns to represent the document embeddings in the form of Gaussian distributions, thereby encoding the uncertainty in its covariance. We propagate the learned uncertainties through linear classifiers that benefit zero-shot cross-lingual topic identification. Our experiments on 17 languages show that the proposed multilingual Bayesian document model performs competitively, when compared to other systems based on large-scale neural networks (LASER, XLM-R, mUSE) on 8 high-resource languages, and outperforms these systems on 9 mid-resource languages. We revisit cross-lingual topic identification in zero-shot settings by taking a deeper dive into current datasets, baseline systems and the languages covered. We identify shortcomings in the existing evaluation protocol (MLDoc dataset), and propose a robust alternative scheme, while also extending the cross-lingual experimental setup to 17 languages. Finally, we consolidate the observations from all our experiments, and discuss points that can potentially benefit the future research works in applications relying on cross-lingual transfers.

cs.CL↗

arXiv2023

Strategies for improving low resource speech to text translation relying on pre-trained ASR models

This paper presents techniques and findings for improving the performance of low-resource speech to text translation (ST). We conducted experiments on both simulated and real-low resource setups, on language pairs English - Portuguese, and Tamasheq - French respectively. Using the encoder-decoder framework for ST, our results show that a multilingual automatic speech recognition system acts as a good initialization under low-resource scenarios. Furthermore, using the CTC as an additional objective for translation during training and decoding helps to reorder the internal representations and improves the final translation. Through our experiments, we try to identify various factors (initializations, objectives, and hyper-parameters) that contribute the most for improvements in low-resource setups. With only 300 hours of pre-training data, our model achieved 7.3 BLEU score on Tamasheq - French data, outperforming prior published works from IWSLT 2022 by 1.6 points.

cs.CL↗