Team QUESPA System 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 QUESPA team submitted three unconstrained speech translation (ST) systems for the Quechua to Spanish track of the IWSLT 2026 Evaluation Campaign. Their best-performing system, "contrastive 2," improved upon previous models by integrating a high-performing pre-trained language model for end-to-end ST, avoiding cascading, and incorporating additional Quechua-Collao text. Fine-tuning Microsoft's SpeechT5 model with targeted data augmentation achieved a BLEU score of 27.2 on the official evaluation set. The team also explored prompt-based machine translation using large language models like Gemini, DeepSeek, GPT-5, Claude, and Qwen for the first time. Additionally, they introduced SIDON, an audio enhancement framework. The analysis provides a comparative view across four IWSLT submissions, detailing the impact of synthetic data, external resources, and audio enhancement on fine-tuning performance, underscoring the combined benefits of PLM-based ST, LLM prompting, and ASR enhancement for low-resource speech translation.

Key takeaway

For NLP Engineers developing low-resource speech translation systems, integrating pre-trained language models like SpeechT5 for end-to-end ST is crucial. You should also explore targeted data augmentation with additional text resources and evaluate prompt-based machine translation using advanced LLMs such as Gemini or GPT-5. Additionally, consider implementing audio enhancement frameworks like SIDON to improve input quality, as these complementary approaches significantly boost overall performance in challenging low-resource scenarios.

Key insights

PLM-based speech translation, LLM prompting, and audio enhancement collectively advance low-resource speech translation performance.

Principles

• End-to-end ST benefits from PLMs.
• Data augmentation boosts fine-tuning.
• Audio enhancement improves ST quality.

Method

Fine-tune SpeechT5 with augmented Quechua-Collao text for end-to-end ST. Evaluate prompt-based MT using Gemini, DeepSeek, GPT-5, Claude, and Qwen. Apply SIDON for audio enhancement.

In practice

• Fine-tune SpeechT5 for ST.
• Augment low-resource text data.
• Experiment with LLM prompting.

Topics

• Speech Translation
• Low-Resource Languages
• Pre-trained Language Models
• LLM Prompting
• Data Augmentation
• Audio Enhancement
• IWSLT Evaluation

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Low-resource Language Discrimination Towards Chinese Dialects with Transfer learning and Data Augmentation — Transfer learning and data augmentation effectively overcome scarce annotation resources for low-resource Chinese dialect discrimination.
• How I Improved Speech-to-Text Accuracy — A two-pass LLM post-processing method effectively corrects common speech-to-text transcription errors.
• Multilingual and Multimodal LLMs in the Wild: Building for Low-Resource Languages — Multimodal LLMs must integrate text, speech, and vision for low-resource languages, moving beyond English-centric benchmarks.
• Leveraging Audio-LLMs to Filter Speech-to-Speech Training Data — An audio-LLM can effectively filter noisy speech-to-speech translation data using a Rank-to-Distill strategy, improving S2ST performance.
• Cross-Lingual Token Arbitrage: Optimizing Code Agent Context Windows via Local LLM Preprocessing — Proactive, edge-side prompt rewriting with a local LLM significantly reduces token costs for coding agents.
• Into the Future: Better Language Support 🌍 — Global language support, especially for low-resource languages, is a critical challenge and key area of excitement in voice technology.

Open in AIssential →

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