AI BriefingRevisiting ASR Error Correction via Dedicated Models
Key point
This work proposes a method that improves the efficiency and accuracy of ASR error correction by leveraging a seq2seq model lighter than an LLM.
Details
Existing ASR (Automatic Speech Recognition) error correction methods have mostly relied on text-only models that fail to recognize ASR's error patterns. Recently, there have been attempts to apply LLMs, but high latency and hallucination issues are pointed out as limitations.
This study addresses this problem through a compact seq2seq model trained on error data extracted from real and synthetic audio. To scale up training, a synthetic corpus was built via a cascaded approach combining TTS and ASR, and it was confirmed that matching the diversity of real error distributions is key.
The newly proposed Correction-first decoding approach has the correction model generate candidates, which are then rescored using the ASR's acoustic scores.
The key results are as follows:
- Efficiency: Achieves superior performance while using 15x fewer parameters compared to LLMs
- Accuracy: Achieves 1.5/3.3% WER on LibriSpeech test-clean/other (outperforming LLMs)
- Generality: Demonstrates generalization performance across various ASR architectures and domains, including CTC, Seq2seq, and Transducer
- Precision: Enables more precise correction in low-error regions, where LLMs are weak
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Wispr Unveils 'Canto', a Real-Time Transcription-Specialized Voice Model Achieving Lower WER Than Competitors via GRPO Training