Year: 2026
Type: conference-paper
Abstract: Culturally grounded speech remains underrepresented in existing Automatic Speech Recognition (ASR) resources for Indian languages.We introduce Konkani Daan, a community-driven initiative to collect culturally representative Goan Konkani speech via a participatory web platform, currently comprising over 43.9 hours of 16 kHz recordings with region-specific metadata.To evaluate real-world language coverage, we test a strong Indian multilingual ASR model, AI4Bharat IndicConformer-600M, in zero-shot mode on the Konkani Daan development set (379 utterances), obtaining a Word Error Rate (WER) of 46.46% and Character Error Rate (CER) of 15.47%, indicating substantial domain and cultural mismatch despite nominal support for Konkani.We additionally evaluate a previously developed Konkani ASR model in a cross-corpus setting and conduct a qualitative error analysis of outputs from both systems, identifying recurring challenges including compound-word segmentation, digit-word normalisation, named-entity distortion, and orthographic variation.These findings highlight the need for culturally informed resource design and normalisation-aware evaluation for low-resource Indian languages.
Authors: Milind Shivolkar, Vaibhav Gawas, Jyoti Pawar
Institution: Goa University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166540756