DziriTalk: An algerian desktop voice assistant
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Faculty of Science
Abstract
Voice User Interfaces (VUIs) are pivotal in modern Human-
Computer Interaction (HCI), yet mainstream assistants consistently
fail to process regional, non-standardized dialects like Algerian
Darija due to their reliance on Modern Standard Arabic or foreign
datasets. This creates a stark digital accessibility barrier for over 40
million speakers.
This thesis presents DziriTalk, an offline, privacy-focused desktop
voice assistant engineered explicitly for the Algerian dialect. To
address low-resource challenges and code-switching, we implement
a dual-model pipeline: a BiGRU text classifier and a CNN-BiLSTM
audio fallback, paired with Whisper-based ASR and a three-tier
TTS. The NLU module maps to a system automation layer, enabling
desktop control via spoken dialect.
Evaluation on a collected Darija dataset (110 speakers, 400+
recordings) yields 94% text and 98% audio accuracy, with under 3
seconds latency on CPU-only hardware. Notably, an SVM baseline
drops from 96.93% on clean text to 63% in the vocal pipeline,
confirming the critical value of end-to-end evaluation. Operating
fully offline, DziriTalk bridges the digital divide, allowing users to
interact naturally without frustration or internet dependency