CATENG Submission for the IWSLT 2026: Dialectal and Low-resource Speech Translation Task

2026-06-29 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

The CATENG team submitted three systems to the IWSLT 2026 Dialectal and Low-Resource Speech Translation shared task for Catalan–English (CA–EN). Addressing challenges posed by Catalan's dialectal diversity and limited speech technology representation, the team evaluated two cascaded approaches and one end-to-end model. Their primary system combined a Mamba-based ASR (ConMamba) with a fine-tuned NLLB-200 MT model. A contrastive system utilized Whisper-v3 for ASR, also paired with NLLB-200, while an end-to-end SpeechT5 model incorporated data augmentation. Experiments on the 15-hour IWSLT 2026 Catalan dataset, supplemented by large-scale parallel text, revealed that cascaded systems generally outperformed end-to-end ST. Notably, the Whisper-v3 + NLLB combination achieved 44.7 BLEU and 65.1 chrF. The findings emphasize that ASR quality, rather than MT capacity, primarily limits performance, and Mamba-based ASR models offer competitive results, underscoring the need for robust speech representations and comprehensive dialectal coverage.

Key takeaway

For Machine Learning Engineers developing speech translation systems for dialectal or low-resource languages, prioritize robust ASR components. Your system's overall performance is more likely constrained by ASR quality than by MT capacity. Consider cascaded architectures, specifically evaluating Mamba-based ASR models like ConMamba or established options like Whisper-v3, paired with fine-tuned MT models such as NLLB-200, to achieve superior results over end-to-end approaches. Focus on acquiring diverse dialectal speech data to enhance ASR representations.

Key insights

Cascaded ASR-MT systems outperform end-to-end speech translation, with ASR quality being the primary performance bottleneck.

Principles

• ASR quality is critical for cascaded speech translation performance.
• Robust speech representations improve dialectal speech translation.
• Cascaded ASR-MT often surpasses end-to-end ST.

Method

Evaluated cascaded ASR+MT (ConMamba/Whisper-v3 with NLLB-200) and end-to-end SpeechT5 models on dialectal Catalan–English, using the IWSLT 2026 dataset and parallel text.

In practice

• Prioritize ASR model selection for speech translation tasks.
• Consider Mamba-based ASR for competitive results.
• Focus on dialectal coverage in speech datasets.

Topics

• Speech Translation
• Dialectal Catalan
• Low-Resource NLP
• Cascaded ASR-MT
• Mamba ASR
• Whisper-v3

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

Related on AIssential

• BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language — Fine-tuning pre-trained ASR models with modest, newly created corpora can establish functional speech recognition for low-resource languages.
• Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild — Hy-MT2 delivers high-quality, efficient multilingual translation with strong instruction-following across diverse real-world and resource-constrained scenarios.
• IBM Releases Two Granite Speech 4.1 2B Models: Autoregressive ASR with Translation and Non-Autoregressive Editing for Fast Inference — IBM's new Granite Speech 4.1 2B models offer both high-accuracy multilingual ASR/translation and extremely fast non-autoregressive inference.
• Can Voice Agents Handle Bilingual Customers? Benchmarking Frontier ASR on Code-Switched Speech — Frontier ASR models handle code-switched speech with minimal performance degradation, but error patterns vary by language pair and model.
• Giving Voice to the Constitution: Low-Resource Text-to-Speech for Quechua and Spanish Using a Bilingual Legal Corpus — Cross-lingual transfer in TTS effectively mitigates data scarcity for low-resource languages like Quechua.
• MLLP-VRAIN UPV system for the IWSLT 2026 Simultaneous Speech Translation task — The system combines Parakeet and Qwen 3.5 with adaptive policies and RAG for improved SimulST quality and latency.

Open in AIssential →

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