arXiv subjects

Sathvik Udupa

Publications and source records attributed to Sathvik Udupa.

13 recordsLinked to original sources

arXiv2026

TurnBench: A Multi-Domain Benchmark for Turn-Taking Dynamics in Spoken Dialogue

Speakers in natural conversation take turns speaking and listening, deciding in real time when to take, hold, or yield the floor. However, turn-taking evaluation remains limited due to the lack of a consistent, linguistically grounded evaluation protocol and hand-annotated data covering diverse conversation types. To address this, we present TurnBench, a multi-domain benchmark that pairs a 30-hour, hand-labeled corpus of dyadic human conversation with a standardized evaluation protocol for end-of-turn and interruption detection. We set conversation type as a controllable experimental variable, covering six distinct interaction styles, and triple-annotate each conversation. Benchmarking 14 heterogeneous turn-taking systems, we find end-of-turn recall stable across types, while interruption false positives are strongly type-dependent and concentrated in backchannel-dense interaction styles. Although in smooth floor transfers human listeners begin speaking a median 151 ms before the current turn ends, no current system performs equivalently without incurring excessive false positives. We release our corpus, a 104-hour training set, and a public leaderboard with an interactive dataset viewer at https://turnbench.sesame.com

eess.AS↗

arXiv2026

Grounding Spoken LLMs in Multi-Speaker Audio via Diarization Conditioning

We propose diarization-conditioned spoken language models (SLMs), a strategy for extending SLMs to far-field multi-talker audio. Rather than adapting the decoder via Serialized Output Training, which risks catastrophic forgetting, we condition the acoustic encoder on diarization masks to extract target-speaker representations, keeping the decoder frozen. We instantiate this as Dixtral, integrating a Diarization Conditioned Whisper (DiCoW) encoder into the Voxtral SLM. On AMI, NOTSOFAR-1, LibriSpeechMix, and Mixer6, Dixtral outperforms Gemini 3.0 Flash, VibeVoice, and Voxtral Mini Transcribe V2 on speaker-attributed transcription by 29.0%, 19.8%, and 16.0% absolute cpWER respectively. On a novel long-form multi-speaker QA benchmark, zero-shot Dixtral matches Gemini on far-field content understanding, and when fine-tuned surpasses both Gemini and Voxtral operating on close-talk across all tasks.

eess.AS↗

arXiv2026

Endpoint Anticipation for Low-Latency Spoken Dialogue

While low-latency interaction is critical for spoken dialogue, cascaded architectures are often bottlenecked by reactive turn-completion detection. We propose Endpoint Anticipation, shifting from reactive detection to proactive forecasting of end-of-turn signals. Our speech-based model anticipates endpoints upto 2.56 seconds in advance, enabling speculative execution of LLM and TTS pipelines on partial context. We introduce metrics to quantify the trade-off between realized latency reduction and computational redundancy. Evaluation across conversational and task-oriented datasets shows our model consistently outperforms competitive VAP-based baselines. Integration with the Unmute framework demonstrates a 505 ms average latency reduction with a 28.4% increase in speculative computation, effectively masking sequential bottlenecks to enable complex reasoning in real-time speech-to-speech interaction.

eess.AS↗

arXiv2026

VAANI: Capturing the language landscape for an inclusive digital India

Voice based technologies have the potential to bridge digital accessibility gaps; however, existing datasets fail to capture the linguistic and regional diversity of Indic languages. We present Project VAANI, a large scale multimodal dataset designed to represent India's linguistic landscape across 165 districts. Speech data is collected using image based prompts to elicit spontaneous responses, while images are curated through a separate pipeline covering diverse themes across regions. The dataset undergoes a rigorous multi stage quality control process, combining automated and manual evaluation to ensure high audio quality and transcription accuracy. We release approximately 289K images, 31,255 hours of speech, and 2,043 hours of transcribed audio spanning 105 languages from 28 states and 3 union territories. Many of these languages are represented at this scale for the first time, making VAANI a foundational resource for inclusive speech technology. The dataset enables the development of robust, multilingual, and multimodal models, and supports research in speech recognition, language understanding, and cross-modal learning for underrepresented languages.

eess.AS↗

arXiv2025

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↗

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

Streaming Endpointer for Spoken Dialogue using Neural Audio Codecs and Label-Delayed Training

Accurate, low-latency endpointing is crucial for effective spoken dialogue systems. While traditional endpointers often rely on spectrum-based audio features, this work proposes real-time speech endpointing for multi-turn dialogues using streaming, low-bitrate Neural Audio Codec (NAC) features, building upon recent advancements in neural audio codecs. To further reduce cutoff errors, we introduce a novel label delay training scheme. At a fixed median latency of 160 ms, our combined NAC and label delay approach achieves significant relative cutoff error reductions: 42.7% for a single-stream endpointer and 37.5% for a two-stream configuration, compared to baseline methods. Finally, we demonstrate efficient integration with a codec-based pretrained speech large language model, improving its median response time by 1200 ms and reducing its cutoff error by 35%.

cs.SD↗

arXiv2025

Discovering phoneme-specific critical articulators through a data-driven approach

We propose an approach for learning critical articulators for phonemes through a machine learning approach. We formulate the learning with three models trained end to end. First, we use Acoustic to Articulatory Inversion (AAI) to predict time-varying speech articulators EMA. We also predict the phoneme-specific weights across articulators for each frame. To avoid overfitting, we also add a dropout layer before the weights prediction layer. Next, we normalize the predicted weights across articulators using min-max normalization for each frame. The normalized weights are multiplied by the ground truth $EMA$ and then we try to predict the phones at each frame. We train this whole setup end to end and use two losses. One loss is for the phone prediction which is the cross entropy loss and the other is for the AAI prediction which is the mean squared error loss. To maintain gradient flow between the phone prediction block and the $EMA$ prediction block, we use straight-through estimation. The goal here is to predict the weights of the articulator at each frame while training the model end to end.

eess.AS↗

arXiv2023

Speaking rate attention-based duration prediction for speed control TTS

With the advent of high-quality speech synthesis, there is a lot of interest in controlling various prosodic attributes of speech. Speaking rate is an essential attribute towards modelling the expressivity of speech. In this work, we propose a novel approach to control the speaking rate for non-autoregressive TTS. We achieve this by conditioning the speaking rate inside the duration predictor, allowing implicit speaking rate control. We show the benefits of this approach by synthesising audio at various speaking rate factors and measuring the quality of speaking rate-controlled synthesised speech. Further, we study the effect of the speaking rate distribution of the training data towards effective rate control. Finally, we fine-tune a baseline pretrained TTS model to obtain speaking rate control TTS. We provide various analyses to showcase the benefits of using this proposed approach, along with objective as well as subjective metrics. We find that the proposed methods have higher subjective scores and lower speaker rate errors across many speaking rate factors over the baseline.

eess.AS↗

arXiv2023

Model Adaptation for ASR in low-resource Indian Languages

Automatic speech recognition (ASR) performance has improved drastically in recent years, mainly enabled by self-supervised learning (SSL) based acoustic models such as wav2vec2 and large-scale multi-lingual training like Whisper. A huge challenge still exists for low-resource languages where the availability of both audio and text is limited. This is further complicated by the presence of multiple dialects like in Indian languages. However, many Indian languages can be grouped into the same families and share the same script and grammatical structure. This is where a lot of adaptation and fine-tuning techniques can be applied to overcome the low-resource nature of the data by utilising well-resourced similar languages. In such scenarios, it is important to understand the extent to which each modality, like acoustics and text, is important in building a reliable ASR. It could be the case that an abundance of acoustic data in a language reduces the need for large text-only corpora. Or, due to the availability of various pretrained acoustic models, the vice-versa could also be true. In this proposed special session, we encourage the community to explore these ideas with the data in two low-resource Indian languages of Bengali and Bhojpuri. These approaches are not limited to Indian languages, the solutions are potentially applicable to various languages spoken around the world.

eess.AS↗

arXiv2022

Improved acoustic-to-articulatory inversion using representations from pretrained self-supervised learning models

In this work, we investigate the effectiveness of pretrained Self-Supervised Learning (SSL) features for learning the mapping for acoustic to articulatory inversion (AAI). Signal processing-based acoustic features such as MFCCs have been predominantly used for the AAI task with deep neural networks. With SSL features working well for various other speech tasks such as speech recognition, emotion classification, etc., we experiment with its efficacy for AAI. We train on SSL features with transformer neural networks-based AAI models of 3 different model complexities and compare its performance with MFCCs in subject-specific (SS), pooled and fine-tuned (FT) configurations with data from 10 subjects, and evaluate with correlation coefficient (CC) score on the unseen sentence test set. We find that acoustic feature reconstruction objective-based SSL features such as TERA and DeCoAR work well for AAI, with SS CCs of these SSL features reaching close to the best FT CCs of MFCC. We also find the results consistent across different model sizes.

eess.AS↗

arXiv2022

Real-Time MRI Video synthesis from time aligned phonemes with sequence-to-sequence networks

Real-Time Magnetic resonance imaging (rtMRI) of the midsagittal plane of the mouth is of interest for speech production research. In this work, we focus on estimating utterance level rtMRI video from the spoken phoneme sequence. We obtain time-aligned phonemes from forced alignment, to obtain frame-level phoneme sequences which are aligned with rtMRI frames. We propose a sequence-to-sequence learning model with a transformer phoneme encoder and convolutional frame decoder. We then modify the learning by using intermediary features obtained from sampling from a pretrained phoneme-conditioned variational autoencoder (CVAE). We train on 8 subjects in a subject-specific manner and demonstrate the performance with a subjective test. We also use an auxiliary task of air tissue boundary (ATB) segmentation to obtain the objective scores on the proposed models. We show that the proposed method is able to generate realistic rtMRI video for unseen utterances, and adding CVAE is beneficial for learning the sequence-to-sequence mapping for subjects where the mapping is hard to learn.

eess.AS↗

arXiv2021

Estimating articulatory movements in speech production with transformer networks

We estimate articulatory movements in speech production from different modalities - acoustics and phonemes. Acoustic-to articulatory inversion (AAI) is a sequence-to-sequence task. On the other hand, phoneme to articulatory (PTA) motion estimation faces a key challenge in reliably aligning the text and the articulatory movements. To address this challenge, we explore the use of a transformer architecture - FastSpeech, with explicit duration modelling to learn hard alignments between the phonemes and articulatory movements. We also train a transformer model on AAI. We use correlation coefficient (CC) and root mean squared error (rMSE) to assess the estimation performance in comparison to existing methods on both tasks. We observe 154%, 11.8% & 4.8% relative improvement in CC with subject-dependent, pooled and fine-tuning strategies, respectively, for PTA estimation. Additionally, on the AAI task, we obtain 1.5%, 3% and 3.1% relative gain in CC on the same setups compared to the state-of-the-art baseline. We further present the computational benefits of having transformer architecture as representation blocks.

eess.AS↗