Nuha-Speech: Establishing Foundations for Arabic Speech-LLMs

While multimodal speech-LLMs advance rapidly across major languages, Arabic remains constrained by a persistent shortage of speech-text training assets. The Nuha-Speech project establishes an end-to-end infrastructure to bridge this gap, introducing an Arabic Speech Question-Answering corpus with over 1.5 million instruction samples. Leveraging Qwen-Omni model variants across scales alongside a dedicated evaluation benchmark, the work provides a grounded reference for developing conversational speech models in resource-limited languages.