Mispronunciation Detection on L2-Arctic-plus (test)
53.8PrecisionASR+LLM Cascade SOTA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ASR+LLM Cascade SOTAAudio Encoder=Wav2vec2 Base, LLM=Llama-3.1-8B2026.01 | 53.8 | 17.8 | 26.8 | 1.1 | |
| GPT-4o-Audio2026.01 | 52.7 | 41.3 | 46.3 | 0.2 | |
| Instruction-Tuned ALMAudio Encoder=Whisper Medium, LLM=Mistral-7B2026.01 | 51.6 | 78.2 | 62.1 | 0 | |
| Instruction-Tuned ALMAudio Encoder=Whisper Medium, LLM=Llama-3.1-8B2026.01 | 51.2 | 78.3 | 61.9 | 0 | |
| Instruction-Tuned ALMAudio Encoder=Whisper Large, LLM=Mistral-7B2026.01 | 50.6 | 81.8 | 62.5 | 0 | |
| Instruction-Tuned ALMAudio Encoder=Whisper Small, LLM=Mistral-7B2026.01 | 50.5 | 65.5 | 57.1 | 0 | |
| Instruction-Tuned ALMAudio Encoder=Whisper Small, LLM=Llama-3.1-8B2026.01 | 49.7 | 68.2 | 57.5 | 0 | |
| Instruction-Tuned ALMAudio Encoder=Whisper Large, LLM=Llama-3.1-8B2026.01 | 48.9 | 87.7 | 62.8 | 0 |