aiqbench analysis
This article details the creation of BaltiVoice, a 16.8-hour speech corpus and fine-tuned Whisper ASR system designed to address the lack of resources for the low-resource Balti language. By adapting OpenAI's Whisper model, the authors demonstrate significant reductions in Word and Character Error Rates compared to zero-shot baselines, highlighting the critical role of model capacity in low-data regimes. The release of this dataset and fine-tuned models on HuggingFace provides a valuable resource for expanding automatic speech recognition capabilities into underrepresented Tibetic languages.
BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language
arXiv:2606.03504v3 Announce Type: replace-cross Abstract: We present BaltiVoice, a 16.8-hour read-speech corpus for Balti (ISO 639-3: bft), a Tibetic language spoken in Gilgit-Baltistan, Pakistan, with no prior publicly available ASR resources. The corpus contains 10,060 validated utterances in native Nastaliq script, derived from Mozilla Common Voice recordings. Fine-tuning OpenAI Whisper-small yields a Word Error Rate (WER) of 24.78% and a Character Error Rate (CER) of 8.30% after training for 5 epochs (3,000 steps) on the 538-utterance speaker-disjoint validation set, down from a zero-shot baseline of 159.19% WER and 152.52% CER. A Whisper-base fine-tuned on the same data achieves 44.54% WER and 15.61% CER, confirming that model capacity matters for this low-resource setting. The dataset, fine-tuned model, and a live transcription demo are publicly available on HuggingFace.
Abstract:We present BaltiVoice, a 16.8-hour read-speech corpus for Balti (ISO 639-3: bft), a Tibetic language spoken in Gilgit-Baltistan, Pakistan, with no prior publicly available ASR resources. The corpus contains 10,060 validated utterances in native Nastaliq script, derived from Mozilla Common Voice recordings. Fine-tuning OpenAI Whisper-small yields a Word Error Rate (WER) of 24.78% and a Character Error Rate (CER) of 8.30% after training for 5 epochs (3,000 steps) on the 538-utterance speaker-disjoint validation set, down from a zero-shot baseline of 159.19% WER and 152.52% CER. A Whisper-base fine-tuned on the same data achieves 44.54% WER and 15.61% CER, confirming that model capacity matters for this low-resource setting. The dataset, fine-tuned model, and a live transcription demo are publicly available on HuggingFace.
6 pages, 3 figures, 4 tables. Code and data available at this https URL
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
I.2.7; I.5.4; J.5
arXiv:2606.03504 [cs.CL]
(or arXiv:2606.03504v3 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.03504 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 234 words
Related coverage
BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language
arXiv · Jun 3, 2026
BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language
arXiv · Jun 11, 2026
A Speech Corpus for Mizo Automatic Speech Recognition: Whisper and SraVaani 1.0 Fine-Tuning with Morphology-Aware Evaluation
arXiv · Aug 21, 2026
Robust Assamese Speech Recognition through Controlled Fine-Tuning of Whisper Models
arXiv · Jul 21, 2026