Polyglot-Lion: Efficient Multilingual ASR for Singapore via Balanced Fine-Tuning of Qwen3-ASR

2026-06-26 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

Polyglot-Lion is a new family of compact multilingual Automatic Speech Recognition (ASR) models developed for Singapore's diverse linguistic landscape, encompassing English, Mandarin, Tamil, and Malay. These models are derived from fine-tuning Qwen3-ASR-0.6B and Qwen3-ASR-1.7B on publicly available speech corpora. A key aspect of their development involves a balanced sampling strategy that ensures an equal number of training utterances per language and intentionally omits language-tag conditioning, enabling the models to implicitly identify languages from audio. On 12 benchmarks, Polyglot-Lion-1.7B achieved an average error rate of 14.85, closely competing with the 6x larger MERaLiON-2-10B-ASR (14.32). Training cost was \$81 on a single RTX PRO 6000 GPU, and inference throughput is approximately 20x faster than MERaLiON, at 0.10 s/sample versus 2.02 s/sample.

Key takeaway

For Machine Learning Engineers developing multilingual ASR systems, consider adopting balanced fine-tuning strategies on moderate-scale pretrained models. You can achieve competitive performance and significantly reduce training costs to around \$81 per model, while also boosting inference throughput by up to 20x compared to larger, more resource-intensive alternatives. This approach allows for deployment-ready solutions tailored to specific linguistic needs, such as Singapore's four languages, without requiring massive computational resources.

Key insights

Balanced fine-tuning of moderate-scale ASR models yields cost-effective, high-performance multilingual speech recognition.

Principles

• Equalize training utterances per language.
• Omit explicit language-tag conditioning.
• Moderate-scale models can outperform larger systems.

Method

Fine-tuning Qwen3-ASR models on public corpora using balanced language sampling and implicit language identification from audio.

In practice

• Deploy compact ASR for multilingual environments.
• Reduce ASR training and inference costs.

Topics

• Multilingual ASR
• Fine-tuning
• Qwen3-ASR
• Speech Recognition
• Model Efficiency
• Singapore Linguistics

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

Related on AIssential

• 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.
• 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.
• Multilingual Fine-Tuning via Localized Gradient Conflict Resolution — Localized gradient conflict resolution via parameter buckets significantly improves multilingual LLM fine-tuning efficiency and performance.
• LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification — Iterative imbalance-aware fine-tuning of LLMs, combined with augmentation and post-processing, significantly improves rare-class NLP performance.
• NIM4-ASR: Towards Efficient, Robust, and Customizable Real-Time LLM-Based ASR — NIM4-ASR optimizes LLM-based ASR for efficiency and robustness by redesigning multi-stage training and integrating production-ready features.
• Towards Truly Multilingual ASR: Generalizing Code-Switching ASR to Unseen Language Pairs — Code-switching ASR generalization to unseen language pairs remains limited despite model merging and domain generalization efforts.

Open in AIssential →

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