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

Dimensions overall score 8.0

• Polyglot-Lion-1.7B achieves an average error rate of 14.85 on 12 benchmarks spanning English, Mandarin, Tamil, and Malay.result 95%
• Polyglot-Lion-1.7B is competitive with MERaLiON-2-10B-ASR (14.32 error rate) while being 6x smaller.result 90%
• Training cost for Polyglot-Lion was $81 on a single RTX PRO 6000 GPU compared to $18,862 for the 128-GPU baseline.technical 95%
• Inference throughput is approximately 20x faster than MERaLiON at 0.10 s/sample versus 2.02 s/sample.technical 90%
• The model uses a balanced sampling strategy that equalizes the number of training utterances per language.method 85%
• The model deliberately omits language-tag conditioning so that it learns to identify languages implicitly from audio.method 85%
• The model is limited to four languages (English, Mandarin, Tamil, Malay) without easy expansion.limitation 80%
• The model relies on publicly available data which may not cover all accents or domains.limitation 80%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card