A data-centric approach to performance improvement in under-resourced ASR: The case of Dënë Sųłıné

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

A study presented at the AmericasNLP 2026 workshop by Olga Kriukova, Olga Lovick, and Antti Arppe investigates data-centric approaches to improve Automatic Speech Recognition (ASR) for Dënë Sųłıné, an under-resourced language. The research explores various strategies to enhance the quality of both audio recordings and transcriptions within mixed-quality datasets. Experiments revealed that reducing non-phonemic spelling variation in the corpus substantially improves ASR model generalization and recognition accuracy. Furthermore, increasing the volume of manually reviewed transcriptions consistently lowered word and character error rates. Conversely, audio enhancement techniques were found to slightly decrease ASR performance, highlighting complex trade-offs inherent in developing ASR for low-resource languages.

Key takeaway

For NLP Engineers developing ASR systems for under-resourced languages, prioritize data quality over raw quantity or generic audio enhancements. Focus your efforts on standardizing transcription spelling variations and investing in manual review of existing transcriptions to significantly improve model accuracy and generalization. Be cautious with audio enhancement, as it may degrade performance in low-resource contexts.

Key insights

Data quality, especially transcription consistency, is crucial for low-resource ASR performance.

Principles

• Minimize non-phonemic spelling variation.
• Prioritize manual transcription review.
• Audio enhancement can be detrimental.

Method

The study systematically evaluates data preparation techniques for ASR, focusing on transcription normalization and manual review to enhance training data quality.

In practice

• Implement transcription normalization tools.
• Allocate resources for manual transcription.
• Re-evaluate audio preprocessing steps.

Topics

• Automatic Speech Recognition
• Low-Resource Languages
• Dënë Sųłıné
• Data-Centric AI
• Transcription Normalization
• Data Quality
• Indigenous Languages NLP

Best for: Research Scientist, AI Scientist, 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.
• 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.
• Improving low-resource ASR using bilingual fine-tuning with language identification: a cross-linguistic evaluation — Bilingual fine-tuning enhances low-resource ASR, particularly when language identification is accurate or explicitly guided during inference.
• Scaling Human and G2P Supervision for Robust Phonetic Transcription — G2P supervision for phonetic transcription has diminishing returns beyond 20-30 hours of human annotation.
• Poster: Exploring the Limits of Audio-Based Detection of Turkish Phone Call Scams — Transcript-based LLM analysis significantly improves Turkish phone scam detection over raw audio, even with uncorrected ASR.
• The Data Problem in Low-Resource Languages — Data availability, quality, and standardization are critical for NLP model performance across diverse languages.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.