Tag
The Itgan team describes a unified Whisper+LoRA recipe adapted for the three NADI 2026 Arabic ASR subtasks, showing that per-dialect specialists, base-model choice, and training-free weight averaging each dominate for different subtasks rather than any single novel component.
Clean English speech is commoditized, but in 2026, ASR faces challenges such as overlapping speech recognition, real-time transcription, dialect and terminology handling, and optimizing the speech tower to avoid burdening Speech LLM.
This paper presents an MLIR-based compilation method for large language models, using two custom dialects (TopOp and TpuOp) to lower models from framework-agnostic semantics to hardware-specific instructions. It also introduces a three-stage static compilation for autoregressive inference stages: prefill, prefill_kv, and decode.