Ajinkya Kulkarni

Hawau Olamide Toyin, Amirbek Djanibekov, Ajinkya Kulkarni et al.

We present ArTST, a pre-trained Arabic text and speech transformer for supporting open-source speech technologies for the Arabic language. The model architecture follows the unified-modal framework, SpeechT5, that was recently released for English, and is focused on Modern Standard Arabic (MSA), with plans to extend the model for dialectal and code-switched Arabic in future editions. We pre-trained the model from scratch on MSA speech and text data, and fine-tuned it for the following tasks: Automatic Speech Recognition (ASR), Text-To-Speech synthesis (TTS), and spoken dialect identification. In our experiments comparing ArTST with SpeechT5, as well as with previously reported results in these tasks, ArTST performs on a par with or exceeding the current state-of-the-art in all three tasks. Moreover, we find that our pre-training is conducive for generalization, which is particularly evident in the low-resource TTS task. The pre-trained model as well as the fine-tuned ASR and TTS models are released for research use.

Atharva Kulkarni, Ajinkya Kulkarni, Miguel Couceiro et al.

Recently, large pre-trained multilingual speech models have shown potential in scaling Automatic Speech Recognition (ASR) to many low-resource languages. Some of these models employ language adapters in their formulation, which helps to improve monolingual performance and avoids some of the drawbacks of multi-lingual modeling on resource-rich languages. However, this formulation restricts the usability of these models on code-switched speech, where two languages are mixed together in the same utterance. In this work, we propose ways to effectively fine-tune such models on code-switched speech, by assimilating information from both language adapters at each language adaptation point in the network. We also model code-switching as a sequence of latent binary sequences that can be used to guide the flow of information from each language adapter at the frame level. The proposed approaches are evaluated on three code-switched datasets encompassing Arabic, Mandarin, and Hindi languages paired with English, showing consistent improvements in code-switching performance with at least 10\% absolute reduction in CER across all test sets.

Ajinkya Kulkarni, Atharva Kulkarni, Sara Abedalmonem Mohammad Shatnawi et al.

At present, Text-to-speech (TTS) systems that are trained with high-quality transcribed speech data using end-to-end neural models can generate speech that is intelligible, natural, and closely resembles human speech. These models are trained with relatively large single-speaker professionally recorded audio, typically extracted from audiobooks. Meanwhile, due to the scarcity of freely available speech corpora of this kind, a larger gap exists in Arabic TTS research and development. Most of the existing freely available Arabic speech corpora are not suitable for TTS training as they contain multi-speaker casual speech with variations in recording conditions and quality, whereas the corpus curated for speech synthesis are generally small in size and not suitable for training state-of-the-art end-to-end models. In a move towards filling this gap in resources, we present a speech corpus for Classical Arabic Text-to-Speech (ClArTTS) to support the development of end-to-end TTS systems for Arabic. The speech is extracted from a LibriVox audiobook, which is then processed, segmented, and manually transcribed and annotated. The final ClArTTS corpus contains about 12 hours of speech from a single male speaker sampled at 40100 kHz. In this paper, we describe the process of corpus creation and provide details of corpus statistics and a comparison with existing resources. Furthermore, we develop two TTS systems based on Grad-TTS and Glow-TTS and illustrate the performance of the resulting systems via subjective and objective evaluations. The corpus will be made publicly available at www.clartts.com for research purposes, along with the baseline TTS systems demo.

Ajinkya Kulkarni, Hanan Aldarmaki

In this paper, we describe a spoken Arabic dialect identification (ADI) model for Arabic that consistently outperforms previously published results on two benchmark datasets: ADI-5 and ADI-17. We explore two architectural variations: ResNet and ECAPA-TDNN, coupled with two types of acoustic features: MFCCs and features exratected from the pre-trained self-supervised model UniSpeech-SAT Large, as well as a fusion of all four variants. We find that individually, ECAPA-TDNN network outperforms ResNet, and models with UniSpeech-SAT features outperform models with MFCCs by a large margin. Furthermore, a fusion of all four variants consistently outperforms individual models. Our best models outperform previously reported results on both datasets, with accuracies of 84.7% and 96.9% on ADI-5 and ADI-17, respectively.

Karl El Hajal, Ajinkya Kulkarni, Enno Hermann et al.

While recent zero-shot multi-speaker text-to-speech (TTS) models achieve impressive results, they typically rely on extensive transcribed speech datasets from numerous speakers and intricate training pipelines. Meanwhile, self-supervised learning (SSL) speech features have emerged as effective intermediate representations for TTS. Further, SSL features from different speakers that are linearly close share phonetic information while maintaining individual speaker identity. In this study, we introduce kNN-TTS, a simple and effective framework for zero-shot multi-speaker TTS using retrieval methods which leverage the linear relationships between SSL features. Objective and subjective evaluations show that our models, trained on transcribed speech from a single speaker only, achieve performance comparable to state-of-the-art models that are trained on significantly larger training datasets. The low training data requirements mean that kNN-TTS is well suited for the development of multi-speaker TTS systems for low-resource domains and languages. We also introduce an interpolation parameter which enables fine-grained voice morphing. Demo samples are available at https://idiap.github.io/knn-tts

Xiuzhan Guo, Arthur Berrill, Ajinkya Kulkarni et al.

Ontology operations, e.g., aligning and merging, were studied and implemented extensively in different settings, such as, categorical operations, relation algebras, typed graph grammars, with different concerns. However, aligning and merging operations in the settings share some generic properties, e.g., idempotence, commutativity, associativity, and representativity, labeled by (I), (C), (A), and (R), respectively, which are defined on an ontology merging system $(\mathfrak{O},\sim,\merge)$, where $\mathfrak{O}$ is a set of the ontologies concerned, $\sim$ is a binary relation on $\mathfrak{O}$ modeling ontology aligning and $\merge$ is a partial binary operation on $\mathfrak{O}$ modeling ontology merging. Given an ontology repository, a finite set $\mathbb{O}\subseteq \mathfrak{O}$, its merging closure $\widehat{\mathbb{O}}$ is the smallest set of ontologies, which contains the repository and is closed with respect to merging. If (I), (C), (A), and (R) are satisfied, then both $\mathfrak{O}$ and $\widehat{\mathbb{O}}$ are partially ordered naturally by merging, $\widehat{\mathbb{O}}$ is finite and can be computed efficiently, including sorting, selecting, and querying some specific elements, e.g., maximal ontologies and minimal ontologies. We also show that the ontology merging system, given by ontology $V$-alignment pairs and pushouts, satisfies the properties: (I), (C), (A), and (R) so that the merging system is partially ordered and the merging closure of a given repository with respect to pushouts can be computed efficiently.

Sandipana Dowerah, Atharva Kulkarni, Ajinkya Kulkarni et al.

Parallel to the development of advanced deepfake audio generation, audio deepfake detection has also seen significant progress. However, a standardized and comprehensive benchmark is still missing. To address this, we introduce Speech DeepFake (DF) Arena, the first comprehensive benchmark for audio deepfake detection. Speech DF Arena provides a toolkit to uniformly evaluate detection systems, currently across 14 diverse datasets and attack scenarios, standardized evaluation metrics and protocols for reproducibility and transparency. It also includes a leaderboard to compare and rank the systems to help researchers and developers enhance their reliability and robustness. We include 14 evaluation sets, 12 state-of-the-art open-source and 3 proprietary detection systems. Our study presents many systems exhibiting high EER in out-of-domain scenarios, highlighting the need for extensive cross-domain evaluation. The leaderboard is hosted on Huggingface1 and a toolkit for reproducing results across the listed datasets is available on GitHub.

Xiuzhan Guo, Arthur Berrill, Ajinkya Kulkarni et al.

Benjelloun et al. \cite{BGSWW} considered the Entity Resolution (ER) problem as the generic process of matching and merging entity records judged to represent the same real world object. They treated the functions for matching and merging entity records as black-boxes and introduced four important properties that enable efficient generic ER algorithms. In this paper, we shall study the properties which match and merge functions share, model matching and merging black-boxes for ER in a partial groupoid, based on the properties that match and merge functions satisfy, and show that a partial groupoid provides another generic setting for ER. The natural partial order on a partial groupoid is defined when the partial groupoid satisfies Idempotence and Catenary associativity. Given a partial order on a partial groupoid, the least upper bound and compatibility ($LU_{pg}$ and $CP_{pg}$) properties are equivalent to Idempotence, Commutativity, Associativity, and Representativity and the partial order must be the natural one we defined when the domain of the partial operation is reflexive. The partiality of a partial groupoid can be reduced using connected components and clique covers of its domain graph, and a noncommutative partial groupoid can be mapped to a commutative one homomorphically if it has the partial idempotent semigroup like structures. In a finitely generated partial groupoid $(P,D,\circ)$ without any conditions required, the ER we concern is the full elements in $P$. If $(P,D,\circ)$ satisfies Idempotence and Catenary associativity, then the ER is the maximal elements in $P$, which are full elements and form the ER defined in \cite{BGSWW}. Furthermore, in the case, since there is a transitive binary order, we consider ER as ``sorting, selecting, and querying the elements in a finitely generated partial groupoid."

Ajinkya Kulkarni, Sandipana Dowerah, Atharva Kulkarni et al.

Self-supervised learning (SSL) underpins modern audio deepfake detection, yet most prior work centers on a single large wav2vec2-XLSR backbone, leaving compact under studied. We present RAPTOR, Representation Aware Pairwise-gated Transformer for Out-of-domain Recognition a controlled study of compact SSL backbones from the HuBERT and WavLM within a unified pairwise-gated fusion detector, evaluated across 14 cross-domain benchmarks. We show that multilingual HuBERT pre-training is the primary driver of cross-domain robustness, enabling 100M models to match larger and commercial systems. Beyond EER, we introduce a test-time augmentation protocol with perturbation-based aleatoric uncertainty to expose calibration differences invisible to standard metrics: WavLM variants exhibit overconfident miscalibration under perturbation, whereas iterative mHuBERT remains stable. These findings indicate that SSL pre-training trajectory, not model scale, drives reliable audio deepfake detection.

Panagiotis Theocharopoulos, Ajinkya Kulkarni, Mathew Magimai. -Doss

Large language models (LLMs) are increasingly considered for use in high-impact workflows, including academic peer review. However, LLMs are vulnerable to document-level hidden prompt injection attacks. In this work, we construct a dataset of approximately 500 real academic papers accepted to ICML and evaluate the effect of embedding hidden adversarial prompts within these documents. Each paper is injected with semantically equivalent instructions in four different languages and reviewed using an LLM. We find that prompt injection induces substantial changes in review scores and accept/reject decisions for English, Japanese, and Chinese injections, while Arabic injections produce little to no effect. These results highlight the susceptibility of LLM-based reviewing systems to document-level prompt injection and reveal notable differences in vulnerability across languages.

Riccardo Scodellaro, Ajinkya Kulkarni, Frauke Alves et al.

Recent successes in image analysis with deep neural networks are achieved almost exclusively with Convolutional Neural Networks (CNNs), typically trained using the backpropagation (BP) algorithm. In a 2022 preprint, Geoffrey Hinton proposed the Forward-Forward (FF) algorithm as a biologically inspired alternative, where positive and negative examples are jointly presented to the network and training is guided by a locally defined goodness function. Here, we extend the FF paradigm to CNNs. We introduce two spatially extended labeling strategies, based on Fourier patterns and morphological transformations, that enable convolutional layers to access label information across all spatial positions. On CIFAR10, we show that deeper FF-trained CNNs can be optimized successfully and that morphology-based labels prevent shortcut solutions on dataset with more complex and fine features. On CIFAR100, carefully designed label sets scale effectively to 100 classes. Class Activation Maps reveal that FF-trained CNNs learn meaningful and complementary features across layers. Together, these results demonstrate that FF training is feasible beyond fully connected networks, provide new insights into its learning dynamics and stability, and highlight its potential for neuromorphic computing and biologically inspired learning.

Ajinkya Kulkarni, Anna Tokareva, Rameez Qureshi et al.

In the field of spoken language understanding, systems like Whisper and Multilingual Massive Speech (MMS) have shown state-of-the-art performances. This study is dedicated to a comprehensive exploration of the Whisper and MMS systems, with a focus on assessing biases in automatic speech recognition (ASR) inherent to casual conversation speech specific to the Portuguese language. Our investigation encompasses various categories, including gender, age, skin tone color, and geo-location. Alongside traditional ASR evaluation metrics such as Word Error Rate (WER), we have incorporated p-value statistical significance for gender bias analysis. Furthermore, we extensively examine the impact of data distribution and empirically show that oversampling techniques alleviate such stereotypical biases. This research represents a pioneering effort in quantifying biases in the Portuguese language context through the application of MMS and Whisper, contributing to a better understanding of ASR systems' performance in multilingual settings.

Karl El Hajal, Enno Hermann, Ajinkya Kulkarni et al.

Automatic speech recognition (ASR) systems are well known to perform poorly on dysarthric speech. Previous works have addressed this by speaking rate modification to reduce the mismatch with typical speech. Unfortunately, these approaches rely on transcribed speech data to estimate speaking rates and phoneme durations, which might not be available for unseen speakers. Therefore, we combine unsupervised rhythm and voice conversion methods based on self-supervised speech representations to map dysarthric to typical speech. We evaluate the outputs with a large ASR model pre-trained on healthy speech without further fine-tuning and find that the proposed rhythm conversion especially improves performance for speakers of the Torgo corpus with more severe cases of dysarthria. Code and audio samples are available at https://idiap.github.io/RnV .

Ajinkya Kulkarni, Sandipana Dowerah, Tanel Alumae et al.

Audio deepfakes are acquiring an unprecedented level of realism with advanced AI. While current research focuses on discerning real speech from spoofed speech, tracing the source system is equally crucial. This work proposes a novel audio source tracing system combining deep metric multi-class N-pair loss with Real Emphasis and Fake Dispersion framework, a Conformer classification network, and ensemble score-embedding fusion. The N-pair loss improves discriminative ability, while Real Emphasis and Fake Dispersion enhance robustness by focusing on differentiating real and fake speech patterns. The Conformer network captures both global and local dependencies in the audio signal, crucial for source tracing. The proposed ensemble score-embedding fusion shows an optimal trade-off between in-domain and out-of-domain source tracing scenarios. We evaluate our method using Frechet Distance and standard metrics, demonstrating superior performance in source tracing over the baseline system.

Ajinkya Kulkarni, Francisco Teixeira, Enno Hermann et al.

Children are one of the most under-represented groups in speech technologies, as well as one of the most vulnerable in terms of privacy. Despite this, anonymization techniques targeting this population have received little attention. In this study, we seek to bridge this gap, and establish a baseline for the use of voice anonymization techniques designed for adult speech when applied to children's voices. Such an evaluation is essential, as children's speech presents a distinct set of challenges when compared to that of adults. This study comprises three children's datasets, six anonymization methods, and objective and subjective utility metrics for evaluation. Our results show that existing systems for adults are still able to protect children's voice privacy, but suffer from much higher utility degradation. In addition, our subjective study displays the challenges of automatic evaluation methods for speech quality in children's speech, highlighting the need for further research.

Ajinkya Kulkarni, Atharva Kulkarni, Miguel Couceiro et al.

In this paper, we present a bias and sustainability focused investigation of Automatic Speech Recognition (ASR) systems, namely Whisper and Massively Multilingual Speech (MMS), which have achieved state-of-the-art (SOTA) performances. Despite their improved performance in controlled settings, there remains a critical gap in understanding their efficacy and equity in real-world scenarios. We analyze ASR biases w.r.t. gender, accent, and age group, as well as their effect on downstream tasks. In addition, we examine the environmental impact of ASR systems, scrutinizing the use of large acoustic models on carbon emission and energy consumption. We also provide insights into our empirical analyses, offering a valuable contribution to the claims surrounding bias and sustainability in ASR systems.

Adham Ibrahim, Shady Shehata, Ajinkya Kulkarni et al.

Speech emotion recognition (SER) is essential for enhancing human-computer interaction in speech-based applications. Despite improvements in specific emotional datasets, there is still a research gap in SER's capability to generalize across real-world situations. In this paper, we investigate approaches to generalize the SER system across different emotion datasets. In particular, we incorporate 11 emotional speech datasets and illustrate a comprehensive benchmark on the SER task. We also address the challenge of imbalanced data distribution using over-sampling methods when combining SER datasets for training. Furthermore, we explore various evaluation protocols for adeptness in the generalization of SER. Building on this, we explore the potential of Whisper for SER, emphasizing the importance of thorough evaluation. Our approach is designed to advance SER technology by integrating speaker-independent methods.