Pronunciation Training Feedback Generation on L2-Arctic-plus Human Evaluation (12 samples)
3.8SR (Suggestion Relevance)Whisper Large + Llama-3.1-8B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Whisper Large + Llama-3.1-8BFramework=Instruction-tuned ALM, ASR Backbone=Whisper Large, LLM Backbone=Llama-3.1-8B2026.01 | 3.8 | 3.81 | 3.73 | |
| GPT-4o-Audio2026.01 | 2.88 | 3.51 | 3.07 | |
| Qwen2-Audio2026.01 | 2.12 | 2.83 | 2.26 | |
| Wav2vec2 Base + Llama-3.1-8BFramework=ASR+LLM cascade, ASR Backbone=Wav2vec2 Base, LLM Backbone=Llama-3.1-8B2026.01 | 1.8 | 2.5 | 1.9 |