BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language

2026-06-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

BaltiVoice introduces a 16.8-hour read-speech corpus for Balti (ISO 639-3: bft), a Tibetic language previously lacking public Automatic Speech Recognition (ASR) resources. This corpus comprises 10,060 validated utterances in native Nastaliq script, sourced from Mozilla Common Voice recordings. Researchers fine-tuned OpenAI Whisper-small on this new dataset, achieving a Word Error Rate (WER) of 30.07% on a 538-utterance validation set. This represents a significant improvement from Whisper-small's zero-shot baseline WER of 182.18% for Balti. The complete dataset, the fine-tuned model, and a live transcription demonstration are all publicly accessible on HuggingFace.

Key takeaway

For NLP Engineers or Machine Learning Engineers tasked with developing ASR systems for low-resource languages, this work demonstrates a viable path. You should consider creating a focused read-speech corpus, even if relatively small (e.g., 16.8 hours), and fine-tuning a pre-trained model like Whisper-small. This approach can yield substantial performance gains, transforming a non-functional zero-shot baseline into a practical system for your target language.

Key insights

Fine-tuning pre-trained ASR models with modest, newly created corpora can establish functional speech recognition for low-resource languages.

Principles

• Publicly available resources enable ASR development for underserved languages.
• Fine-tuning significantly reduces ASR error rates on new languages.

Method

A 16.8-hour read-speech corpus was created from Mozilla Common Voice, then used to fine-tune OpenAI Whisper-small, achieving a 30.07% WER.

In practice

• Utilize Mozilla Common Voice for low-resource language corpus generation.
• Fine-tune Whisper-small for initial ASR system development.

Topics

• Balti Language
• Automatic Speech Recognition
• Speech Corpus
• Whisper Model
• Low-Resource Languages
• Fine-tuning
• Nastaliq Script

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages — Data scarcity, not phonological complexity, primarily drives ASR errors in low-resource, typologically complex languages.
• Software for Automatic Speech Recognition via Whisper models applied to Oral History interviews in the Portuguese language — Ethos AT provides private, accessible automatic speech transcription for Portuguese using Whisper models.
• Pushing the Limits of On-Device Streaming ASR: A Compact, High-Accuracy English Model for Low-Latency Inference — On-device ASR can achieve high accuracy and low latency on CPU through systematic benchmarking and quantization.
• WAXAL: A large-scale open resource for African language speech technology — WAXAL provides a large, open-access dataset for 27 African languages, fostering inclusive speech technology development.
• PolySpeech-100: A Large-Scale Benchmark for Speech Understanding Across 100+ Languages and Dialects — PolySpeech-100 reveals E2E Speech-LLMs excel on dialects, but open-source models struggle with low-resource languages, and CoT prompting can degrade performance.
• Combining Real and Synthetic Speech for ASR Adaptation in Brazilian Portuguese — Domain-guided synthetic speech effectively augments real data for ASR adaptation in specialized, low-resource scenarios.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.