This repository contains open-sourced refined labels for Thai automatic speech recognition (ASR) datasets. These labels were generated using a self-evolving data refinement framework that iteratively cleans a hybrid corpus of weak and precise labels, leading to significant improvements in ASR performance.
Low-resource languages like Thai often suffer from limited high-quality annotations and noisy, weakly-labeled data. Our approach refines a hybrid dataset—comprising 500 hours of precise labels and 16k hours of weak labels—using a pre-trained Zipformer model combined with a CER-based filtering mechanism. As a result, the refined labels achieve a 16.63% relative reduction in Character Error Rate (CER) on the CommonVoice Thai dataset and provide the community with 14k hours of curated Thai speech data.
This repository supports the TASC-LLM framework, which integrates:
- Self-evolving Data Refinement: Iteratively updates labels by retaining only high-confidence predictions (CER ≤ 10%).
- LLM-enhanced ASR Architecture: Projects refined speech features into a Thai language model space.
- Threshold-aware Sequence Compression: Dynamically prunes redundant frames to reduce computational cost without sacrificing performance.
labels/: Contains the refined label files (refined gigaspeech2 and MSR-86K labels).README.md: This file.