Choosing an ASR model for Dënë Sųłıné: Navigating polysynthesis and unstandardized orthography

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Linguistics · Depth: Advanced, quick

Summary

A study comparing Wav2Vec2 and Whisper ASR architectures on a Dënë Sųłıné dataset reveals that Whisper significantly outperforms Wav2Vec2, contrary to prior research suggesting Wav2Vec2's superiority in low-resource environments. Dënë Sųłıné, an endangered language, presents challenges common to under-resourced languages, including unstandardized orthography, pronunciation variation, and complex polysynthetic morphosyntactic structures distinct from major languages used in pre-training data. While Wav2Vec2 has been reported to excel in such settings (e.g., Coto-Solano et al., 2024; Nahabwe et al., 2025; Williams et al., 2023), this research indicates Whisper's better adaptability to datasets with inconsistent spelling and pronunciation. These findings suggest that ASR model performance is influenced by dataset-specific characteristics, not just architecture, dataset size, or language typology, necessitating further verification across similar inconsistent datasets.

Key takeaway

For NLP Engineers and Research Scientists developing ASR systems for endangered or under-resourced languages, particularly those with unstandardized orthography and pronunciation variation, you should re-evaluate assumptions about model performance. Despite reports favoring Wav2Vec2 in low-resource settings, consider prioritizing Whisper. Your specific dataset's characteristics, such as inconsistent spelling, may lead Whisper to yield significantly better results. Always conduct comparative testing with your target language's unique data challenges.

Key insights

Whisper significantly outperforms Wav2Vec2 on Dënë Sųłıné, adapting better to inconsistent orthography and pronunciation, challenging general low-resource ASR assumptions.

Principles

• Model performance depends on dataset-specific characteristics.
• Inconsistent orthography impacts ASR model selection.
• Typological features alone do not predict ASR success.

Method

The study compared Wav2Vec2 and Whisper ASR architectures by fine-tuning them on a Dënë Sųłıné dataset, evaluating performance against challenges like unstandardized orthography and pronunciation variation.

In practice

• Prioritize Whisper for languages with unstandardized orthography.
• Evaluate ASR models on dataset-specific inconsistencies.
• Do not assume general low-resource ASR model superiority.

Topics

• Automatic Speech Recognition
• Low-Resource Languages
• Wav2Vec2
• Whisper
• Dënë Sųłıné
• Orthography Standardization

Best for: AI Scientist, NLP Engineer, Research Scientist

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.
• Subtitle-Aligned Fine-Tuning of Whisper for Swiss German ASR: Benchmark Contamination, Convention Mismatch, and an Honest Baseline at 25.6% WER (13.8% cWER) — Benchmark contamination significantly inflates reported ASR performance, especially for dialectal variations.
• MURMUR: An Efficient Inference System for Long-Form ASR — Murmur efficiently balances long-form ASR accuracy and latency using a two-level inference system.
• A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition — Fine-tuning pretrained Transformers with specific speech representations and label formats significantly improves Quranic ASR accuracy and efficiency.
• 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.
• Analyzing Error Propagation in Korean Spoken QA with ASR-LLM Cascades — ASR-stage information loss, especially single-character errors in Korean, is the primary bottleneck for ASR-LLM cascade performance.

Open in AIssential →

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