Task-Lens: Cross-Task Utility Based Speech Dataset Profiling for Low-Resource Indian Languages

Swati Sharma; Divya V. Sharma; Anubha Gupta

cs.CLFeb 16, 2026

Task-Lens: Cross-Task Utility Based Speech Dataset Profiling for Low-Resource Indian Languages

Authors: Swati Sharma, Divya V. Sharma, Anubha Gupta

Organizations: SBI Lab, IIIT Delhi

Abstract

The rising demand for inclusive speech technologies amplifies the need for multilingual datasets for Natural Language Processing (NLP) research. However, limited awareness of existing task-specific resources in low-resource languages hinders research. This challenge is especially acute in linguistically diverse countries, such as India. Cross-task profiling of existing Indian speech datasets can alleviate the data scarcity challenge. This involves investigating the utility of datasets across multiple downstream tasks rather than focusing on a single task. Prior surveys typically catalogue datasets for a single task, leaving comprehensive cross-task profiling as an open opportunity. Therefore, we propose Task-Lens, a cross-task survey that assesses the readiness of 50 Indian speech datasets spanning 26 languages for nine downstream speech tasks. First, we analyze which datasets contain metadata and properties suitable for specific tasks. Next, we propose task-aligned enhancements to unlock datasets to their full downstream potential. Finally, we identify tasks and Indian languages that are critically underserved by current resources. Our findings reveal that many Indian speech datasets contain untapped metadata that can support multiple downstream tasks. By uncovering cross-task linkages and gaps, Task-Lens enables researchers to explore the broader applicability of existing datasets and to prioritize dataset creation for underserved tasks and languages.

Figures & tables

Figure 1: Task-Lens: It involves dataset discovery, dataset filtering, feature extraction, followed by utility mapping that aligns dataset features with task needs via a Task-feature relevance matrix labeled as Required and Optional or Not Applicable. A dataset is ‘Task-Ready’ for a task if it satisfies all ’Required’ features for a task. Supported tasks include Automatic Speech Recognition (Monolingual) [T1], Automatic Speech Recognition (Multilingual) [T2], Language Identification [T3], Speaker Verification/Identification [T4], Audio Deepfake Detection [T5], Speech Emotion Recognition [T6], Text-to-Speech (Monolingual) [T7], Text-to-Speech (Multilingual) [T8] and Gender Recognition [T9].

ID	Dataset
$D_{1}$	AccentDB ( Ahamad et al., 2020 )
$D_{2}$	Assamese TTS Corpus ( Tamim et al., 2025 )
$D_{3}$	BhasaAnuvaad ( Jain et al., 2024 )
$D_{4}$	Open-source TTS Voices ( Sodimana et al., 2018 )
$D_{5}$	Bengali Numbers Corpus ( Nahid et al., 2018 )
$D_{6}$	South Asian Crowdsourced Speech

Table 1: List of the 50 Indian speech datasets included in our analysis. Each dataset is assigned a unique identifier (

D_{1}

–

D_{50}

), which is used consistently throughout the paper. The table includes established and lesser-known resources, along with relevant citations.

Feature	Rationale
$f_{1}$ : Audio +	Ensures audio fidelity
Sampling Rate	and robustness
$f_{2}$ : Transcript Presence	Enables text-speech alignment
$f_{3}$ : License Openness	Governs accessibility reuse
$f_{4}$ : Language ID	Supports multilingual tasks
(Cardinality)

Table 2: Utility feature summary.

Feature ( $f_{i}$ )	MO-ASR/	ML-ASR/	LID	SV/SID	ADD	SER	MO-TTS	ML-TTS	GRE
	STT	STT
$f_{1}$ : Audio +	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Sampling Rate
$f_{2}$ : Transcript Presence	$\checkmark$	$\checkmark$	?	?	?	?	$\checkmark$	$\checkmark$	?
$f_{3}$ : License Openness	?	?	?	?	?	?	?	?	?
$f_{4}$ : Language ID	?	$\checkmark$	$\checkmark$	?	?	?	?	$\checkmark$	?

Table 3: Task–feature relevance matrix for dataset–task screening. Symbols:

\checkmark

= Required, ? = Optional / Not Required. For each task, entries marked

\checkmark

form the minimal, exhaustive set of required features under standard task definitions.

Dataset	$T_{1}$	$T_{2}$	$T_{3}$	$T_{4}$	$T_{5}$	$T_{6}$	$T_{7}$	$T_{8}$	$T_{9}$
$D_{1}$	$f_{2}$	$f_{2}$ , $f_{5}$	✓	✓	$f_{7}$	$f_{8}$	$f_{2}$	$f_{2}$ , $f_{5}$	✓
$D_{2}$	✓	$f_{5}$	✓	$f_{6}$	$f_{7}$	$f_{8}$	✓	$f_{5}$	✓
$D_{3}$	✓	✓	✓	$f_{6}$	$f_{7}$	$f_{8}$	✓	✓	✓
$D_{4}$	✓	✓	✓	✓	$f_{7}$	$f_{8}$	✓	✓	✓
$D_{5}$	✓	$f_{5}$	✓	$f_{6}$	$f_{7}$	$f_{8}$	✓	$f_{5}$	$f_{9}$
$D_{6}$	✓	✓	✓	✓	$f_{7}$	$f_{8}$	✓	✓	✓

Table 4: Dataset–task readiness summary. A ✓ denotes Task-Ready (all required features present); missing required features are listed as

f_{i},f_{j}

. Tasks:

T_{1}

=MO-ASR/STT,

T_{2}

=ML-ASR/STT,

T_{3}

=LID,

T_{4}

=SV/SID,

T_{5}

=ADD,

T_{6}

=SER,

T_{7}

=MO-TTS,

T_{8}

=ML-TTS,

T_{9}

=GRE. Notes: Datasets

D_{13}

D_{14}

, and

D_{20}

are excluded because they are music-oriented rather than speech. Dataset

D_{12}

is a challenge release with extensive hours but insufficient public metadata for profiling, so it is omitted here. Although dataset

D_{34}

D_{48}

, and

D_{49}

includes

f_{6}

, which helps map a real speaker’s voice to its synthetic counterpart, the dataset is better suited for anti-spoofing tasks than for speaker verification, where real and synthetic voices are matched.

Figure 2: Distribution of total dataset duration for each task in hours for direct comparison. There is an urgent need of datasets for tasks

T_{4}

(SV/SID),

T_{5}

(ADD), and

T_{6}

(SER).

Language	$T_{1}$	$T_{2}$	$T_{3}$	$T_{4}$	$T_{5}$	$T_{6}$	$T_{7}$	$T_{8}$	$T_{9}$	Datasets
$L_{1}$ : Assamese	537	492	806	465	264	17	537	492	806	$D_{i\in\{2,3,8,15,18,29,35,39,46\}}$
$L_{2}$ : Bengali	1990	1015	2297	635	1418	28	1990	1015	2280	$D_{i\in\{3\!-\!6,8,15\!-\!18,29,34\!-\!36,39,41\!-\!44\}}$
										$D_{i\in\{49,50\}}$
$L_{3}$ : Bhojpuri	120	120	120	0	0	0	120	120	120	$D_{i\in\{9\}}$
$L_{4}$ : Bodo	483	483	550	487	63	0	483	483	550	$D_{i\in\{15,18,35,39,46\}}$
$L_{5}$ : Dogri	200	200	231	204	26	0	200	200	231	$D_{i\in\{15,18,35,39,46\}}$

Table 5: Language-wise Duration Distribution with Dataset Mapping (hours). The tasks with the highest and lowest durations for each language are highlighted in green and red, respectively.

Figure 3: Total speech duration for each Indian language (L 1 –L 26 ) across all 50 datasets. Language L 8 (Hindi) and L 9 (Indian English) have 3,981 and 16,154 hours of data and were excluded from the figure due to duration; they would have dominated the visualization and obscured relative differences among datasets. Languages like

L_{2}

L_{24}

, and

L_{25}

have the highest duration, whereas languages like

L_{23}

L_{3}

L_{11}

, and

L_{12}

are virtually absent.

Low-Resource Language Processing

Low-resource language processing builds and evaluates language technologies for languages with limited data, tools, or digital representation.

See all 565 papers

Apr 20, 2026

BhashaSutra: A Task-Centric Unified Survey of Indian NLP Datasets, Corpora, and Resources

India's linguistic landscape, spanning 22 scheduled languages and hundreds of marginalized dialects, has driven rapid growth in NLP datasets, benchmarks, and pretrained models. However, no dedicated survey consolidates resources developed specifically for Indian languages. Existing reviews either focus on a few high-resource languages or subsume Indian languages within broader multilingual settings, limiting coverage of low-resource and culturally diverse varieties. To address this gap, we present the first unified survey of Indian NLP resources, covering 200+ datasets, 50+ benchmarks, and 100+ models, tools, and systems across text, speech, multimodal, and culturally grounded tasks. We organize resources by linguistic phenomena, domains, and modalities; analyze trends in annotation, evaluation, and model design; and identify persistent challenges such as data sparsity, uneven language coverage, script diversity, and limited cultural and domain generalization. This survey offers a consolidated foundation for equitable, culturally grounded, and scalable NLP research in the Indian linguistic ecosystem.

Raghvendra Kumar, Devankar Raj, Sriparna Saha

Department of Computer Science and Engineering, Indian Institute of Technology Patna, India · Indian Institute of Technology Patna, India

Low-Resource Language Processing

Ask PDF

Jul 26, 2026

Indic DiarBench: A Multilingual Joint Diarization and ASR Benchmark for Indian Languages

In this work, we introduce Indic DiarBench, a speaker diarization and ASR benchmark dataset spanning all 22 scheduled languages of India. This corpus comprises approximately 108 hours of natural multi-speaker audio from near-field meetings, far-field recordings, and in-the-wild audios. All annotations are human-corrected with time-aligned speaker attributed transcriptions. The dataset captures conversational nuance prevalent in Indian speech, such as English code-mixing, dialectal variation, and frequent speaker overlap. To establish a baseline for joint ASR and diarization capabilities we evaluate leading systems including commercial speech APIs and multimodal large language models. Indic DiarBench is released as an open-access resource to advance inclusive, multilingual speech technology research for Indian languages.

Deovrat Mehendale, Aditya Mehndiratta, Dhruv Rathi +2

Sarvam AI, India · AI4Bharat, IIT Madras, India

ASR Evaluation Speech Processing Speaker Diarization Multilingual ASR

Ask PDF

Apr 21, 2026

Voice of India: A Large-Scale Benchmark for Real-World Speech Recognition in India

Existing Indic ASR benchmarks often use scripted, clean speech and leaderboard driven evaluation that encourages dataset specific overfitting. In addition, strict single reference WER penalizes natural spelling variation in Indian languages, including non standardized spellings of code-mixed English origin words. To address these limitations, we introduce Voice of India, a closed source benchmark built from unscripted telephonic conversations covering 15 major Indian languages across 139 regional clusters. The dataset contains 306230 utterances, totaling 536 hours of speech from 36691 speakers with transcripts accounting for spelling variations. We also analyze performance geographically at the district level, revealing disparities. Finally, we provide detailed analysis across factors such as audio quality, speaking rate, gender, and device type, highlighting where current ASR systems struggle and offering insights for improving real world Indic ASR systems.

Kaushal Bhogale, Manas Dhir, Amritansh Walecha +11

Indian Institute of Technology, Madras, India · Josh Talks, India

ASR Evaluation Speech Recognition Benchmarks Automatic Speech Recognition Multilingual ASR

Ask PDF

Abstract

Low-Resource Language Processing

Explore similar work

BhashaSutra: A Task-Centric Unified Survey of Indian NLP Datasets, Corpora, and Resources

Indic DiarBench: A Multilingual Joint Diarization and ASR Benchmark for Indian Languages

Voice of India: A Large-Scale Benchmark for Real-World Speech Recognition in India