Heili: A Local-First Algerian Native Dialect (Darija) Desktop AI Voice Companion and Agent

Heili is a multimodal desktop AI assistant de-signed to operate natively in Algerian native dialect (Darija), a low-resource Arabic dialect with minimal prior NLP coverage. The system runs entirely on a commodity GPU (NVIDIA RTX 4060, 8 GB VRAM) without cloud depen-dency, serving as both a conversational agent and a tool-use system for the Algerian franco-phone user. This paper presents the full design, training pipeline, and evaluation of Heili v8, in-cluding three custom-trained speech components: Moulsot, a Qwen3-ASR-1.7B fine-tune achiev-ing 16.7% word error rate on Algerian speech; an XTTS v2 Darija fine-tune trained on the col-lected corpus that reaches mean opinion score4.2/5.0 on a 50-listener Algerian panel; and an OpenWakeWord model trained on 16,150 anno-tated clips that achieves 0.38 false positives per hour in an AND-gate configuration with Whis-per. A 99-hour Algerian native dialect (Darija) corpus is assembled from 716 YouTube videos across three channels and processed through a seven-stage quality pipeline, yielding 15,567 us-able segments. The FastAPI orchestration layer manages dynamic VRAM allocation so that the full model stack fits within 8 GB through timed eviction. Evaluations demonstrate 82% agent task completion, 89% vision accuracy, and 97% memory retrieval recall. A comparative analysis against ten commercial and open-source assis-tants shows Heili leads on Darija coverage while remaining competitive on response latency (mean 1.77 s end-to-end).Keywords: Algerian native dialect (Darija), low-resource NLP, automatic speech recognition, text-to-speech synthesis, desktop AI assistant, local inference, multimodal