Mapudungun-Spanish Speech Translation: A Low-Resource End-to-End System for the IWSLT 2026 Shared Task

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

Summary

An end-to-end speech translation system for Mapudungun–Spanish was developed for the IWSLT 2026 low-resource task. This system builds upon the Canary-1B-v2 model, employing parameter-efficient fine-tuning with a lightweight adapter. It utilizes an English-centered configuration as a proxy to facilitate translation between the low-resource Indigenous language and Spanish. Experiments demonstrated that the system effectively captures key phonetic patterns, even with limited training data. However, the system also exhibited biases, frequently producing repetitive Spanish outputs. The findings underscore both the potential and the inherent difficulties in adapting large multilingual foundation models for speech translation involving low-resource Indigenous languages like Mapudungun.

Key takeaway

For NLP Engineers developing speech translation systems for low-resource languages, this research indicates that adapting large multilingual models like Canary-1B-v2 via parameter-efficient fine-tuning is a viable approach. You should anticipate challenges such as repetitive outputs and carefully evaluate the trade-offs between data scarcity and translation quality. Consider using English as an intermediate language to bridge translation gaps for under-resourced language pairs.

Key insights

Adapting multilingual foundation models for low-resource Indigenous speech translation is feasible but challenging.

Principles

• Parameter-efficient fine-tuning aids low-resource adaptation.
• English-centered proxies can enable cross-lingual transfer.
• Limited data can lead to repetitive output biases.

Method

The system uses parameter-efficient fine-tuning with a lightweight adapter on Canary-1B-v2, using an English-centered configuration for Mapudungun–Spanish translation.

In practice

• Apply PEFT for low-resource language model adaptation.
• Consider English as a pivot for non-English language pairs.
• Monitor for repetitive outputs in low-resource translation.

Topics

• Speech Translation
• Low-Resource Languages
• Mapudungun
• Canary-1B-v2
• Parameter-Efficient Fine-Tuning
• Multilingual Foundation Models

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• 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.
• Ancient Greek to Modern Greek Machine Translation: A Novel Benchmark and Fine-Tuning Experiments on LLMs and NMT Models — A new AG-MG parallel corpus and benchmark significantly advance Ancient Greek to Modern Greek machine translation.
• Multilingual Reasoning Cascades Need More Context — Multilingual reasoning cascades improve by preserving original query context for final translation.
• 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.
• KIT's Submission to Cross-Lingual Voice Cloning in IWSLT 2026 — Explicit language tags and lexical matching significantly improve cross-lingual voice cloning performance and reduce accent leakage.
• Streaming Speech-to-Text Translation with a SpeechLLM — A novel SpeechLLM adaptively decides when to emit translation tokens, achieving low latency and high quality.

Open in AIssential →

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