Researchers have developed a system for the MLC-SLM 2026 Challenge that adapts the Qwen3-ASR-1.7B model for multilingual, two-speaker conversational speech. The system integrates a speaker diarization front end with the adapted ASR model, processing voice activity, speaker embeddings, and audio segmentation. Adaptation techniques include supervised fine-tuning, LoRA fine-tuning with synthetic speech, and GRPO reinforcement learning, which collectively reduced the word error rate by 6.83 points on the development set and achieved 17.97 tcpMER on the evaluation set. AI
IMPACT Improves accuracy for multilingual conversational AI systems handling multiple speakers.
RANK_REASON The cluster contains an academic paper detailing a new system and adaptation techniques for speech recognition.
- CAMPPlus
- Character Error Rate
- GRPO
- MLC-SLM 2026 Challenge
- Qwen3-ASR-1.7B
- Qwen-ASR-1.7B
- Text To Speech
- word error rate
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
