Nuha-Speech: Building General-Purpose Arabic Speech-LLMs
Nuha-Speech pairs a large Arabic speech QA corpus with fine-tuned Qwen-Omni variants and a dedicated evaluation setup.
The authors say Arabic is still underrepresented in multilingual speech-LLMs. They built an Arabic Speech Question-Answering corpus with more than 1.5 million training samples for instruction tuning across core speech tasks. They then used it to fine-tune Qwen-Omni model variants at different scales and designed task-specific evaluation metrics. ArXiv · AI/CL/LG's note
The authors say Arabic is still underrepresented in multilingual speech-LLMs. They built an Arabic Speech Question-Answering corpus with more than 1.5 million training samples for instruction tuning across core speech tasks. They then used it to fine-tune Qwen-Omni model variants at different scales and designed task-specific evaluation metrics. ArXiv · AI/CL/LG's note
score 5