Preprint
Mixture of LoRA Experts with Multi-Modal and Multi-Granularity LLM Generative Error Correction for Accented Speech Recognition
Bingshen Mu et al., arXiv, 2025
Read the originalRead on October 19, 2026
- What it is
- A 2025 preprint that fine-tunes a large language model to correct speech recognition output, with one LoRA expert per accent combined in a mixture. Its results table reports word error rates for Whisper-large-v3, two audio-language models and its own systems on a combined English test set of nine accents.
- How much weight it carries
- A preprint, not yet peer-reviewed. The nine-accent test set is the authors' own, so the figures compare only the systems in their table, run by the authors. The average column is the authors' figure, and the paper's headline 67.35 percent relative reduction is computed from the average column.
Facts taken from this source
- On the authors' combined nine-accent English test set, FireRedASR-LLM-L had an average word error rate of 7.00 percent.
“FireRedASR-LLM-L [ 63 ] - 8509.09 22.37 7.00 6.74 4.13 5.11 4.81 11.11 13.13 2.62 7.36 17.05” (Table II)
- On the authors' combined nine-accent English test set, Whisper-large-v3 fully fine-tuned on the training data had an average word error rate of 3.29 percent.
“Fully fine-tuned Whisper 1541.62 1541.62 8.29 3.29 3.31 1.93 4.46 3.72 4.77 3.40 1.43 3.60 11.28” (Table II)
- On the authors' combined nine-accent English test set, the HDMoLE mixture of LoRA experts alone had an average word error rate of 2.49 percent.
“HDMoLE 6.72 4329.10 13.70 2.49 2.21 1.33 3.90 1.48 3.35 3.34 0.80 2.60 9.00” (Table II)
- On the authors' combined nine-accent English test set, the HDMoLE system with multi-granularity generative error correction had an average word error rate of 2.07 percent.
“AVG AU CA CN DE IN JP UK US ZA ... + Multi-granularity GER 167.95 8198.03 15.35 2.07 1.22 0.92 3.44 1.11 2.91 2.71 0.54 2.23 8.62” (Table II)
- On the authors' combined nine-accent English test set, Qwen2-Audio had an average word error rate of 5.81 percent.
“Qwen2-Audio [ 62 ] - 8397.09 21.85 5.81 5.35 3.82 5.22 3.89 7.88 10.89 2.21 6.12 10.54” (Table II)
- On the authors' combined nine-accent English test set, the vanilla Whisper-large-v3 baseline had an average word error rate of 6.34 percent.
“Vanilla Whisper [ 41 ] - 1541.62 8.29 6.34 6.08 3.37 6.04 5.56 7.19 14.99 2.02 6.84 13.79” (Table II)