🌍 Can Large Language Models Translate Spoken-Only Languages through International Phonetic Transcription?

[ 🤗Huggingface ]

coming soon

This work introduces two key contributions to address the challenges of applying LLMs to spoken-only languages:

1. UNILANG Framework: A novel approach that enables LLMs to translate spoken-only languages using IPA as an intermediate representation.
2. SOLAN Dataset: The first large-scale bilingual dataset featuring a spoken-only language, Bai, with aligned Chinese and English translations.

Please note that the IPA used in this paper does not follow the standard IPA format.

🧠 UNILANG Framework

UNILANG is the first framework designed to help LLMs understand and translate spoken-only languages by leveraging automatic dictionary construction and knowledge retrieval.

Backbone

• Qwen2.5
• Llama3
• Gemma3

📚 SOLAN Dataset

SOLAN is a curated dataset for spoken-only language translation. It features the Bai language (ISO 639-3: bfs), along with professionally aligned Chinese and English translations.

🎙️ Data Collection

We developed the SOLAN App, a custom Flutter application, to collect high-quality audio samples from native Bai speakers through in-person interviews.

🛠️ Data Processing

Each audio sample is carefully segmented, transcribed using IPA, and validated by linguistic experts to ensure it is suitable for both linguistic analysis and machine learning.

🚀 Requirements

To run the evaluation and processing scripts:

pip install sacrebleu