# Loasr Bench Eval

> Evaluates large speech language models on low-resource automatic speech recognition across 25 languages from 9 typologically diverse families. It probes cross-linguistic generalization, script bias (Latin vs. non-Latin), model scaling effects, and the impact of language-aware prompting on transcription accuracy. Use when the user wants to benchmark on LoASR-Bench, or asks about evaluating this task. Reports error rates.

- Skill: `qhjqhj00/loasr-bench-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/loasr-bench-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/loasr-bench-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/loasr-bench-eval

---


# loasr-bench-eval

> LoASR-Bench: Evaluating Large Speech Language Models on Low-Resource Automatic Speech Recognition Across Language Families — Chen et al. (2026) (arXiv:2603.20042, 2026)

## What this evaluates

Evaluates large speech language models on low-resource automatic speech recognition across 25 languages from 9 typologically diverse families. It probes cross-linguistic generalization, script bias (Latin vs. non-Latin), model scaling effects, and the impact of language-aware prompting on transcription accuracy.

## Datasets

- **LoASR-Bench** — total ?; splits: test (-1)

## Metrics

- `error rates` **(primary)** — range: [0, 1]
  - Standard ASR error rate calculated as (Substitutions + Deletions + Insertions) / Total Characters (or Words), reported as a decimal between 0 and 1. Lower values indicate better transcription accuracy.

## Input / output format

**Input**: Raw audio recording of spoken language, optionally accompanied by a text instruction specifying the target language.

**Output**: Text transcription of the spoken audio.

## Scoring recipe

```python
def compute_error_rate(predictions, references):
    total_errors = 0
    total_chars = 0
    for pred, ref in zip(predictions, references):
        total_errors += levenshtein_distance(pred, ref)
        total_chars += len(ref)
    return total_errors / total_chars if total_chars > 0 else 0.0
```

## Common pitfalls

- Assuming model size scaling linearly improves performance; the benchmark shows diminishing returns (0.04B to 30B only drops error rate from ~0.45 to ~0.32).
- Assuming language-aware prompting always improves results; explicit language names help most languages but can degrade performance on others (e.g., Tamil).
- Overlooking script bias; Latin-script languages consistently show lower error rates than non-Latin scripts across all models.

## Evidence (verbatim from paper)

> Table[I] presents averaged error rates after each language family, revealing notable variations. Romance languages show relatively low error rates across all models, particularly with Qwen3-Omni for most languages. Dravidian languages, however, present significant challenges for the Qwen2-Audio base model but show dramatic improvements with fine-tuning. Furthermore, almost all language families can benefit from language-specific fine-tuning. These findings suggest that lower-resource language families like Dravidian and Indo-Aryan require language-specific fine-tuning and larger multilingual models to achieve comparable performance.

## Citation

```bibtex
@misc{chen2026loasrbench,
  title={LoASR-Bench: Evaluating Large Speech Language Models on Low-Resource Automatic Speech Recognition Across Language Families},
  author={Chen et al. (2026)},
  year={2026},
  note={arXiv:2603.20042}
}
```

- arXiv: 2603.20042