DZIRI VOICEBOT: AN END-TO-END LOW-RESOURCE SPEECH-TO-SPEECH CONVERSATIONAL SYSTEM FOR ALGERIAN DIALECT

This paper addresses the significant lack of speech-based conversational AI for the Algerian dialect (Darija). Because most modern speech technologies are optimized for high-resource languages, they struggle with the unique characteristics of Algerian Darija, such as the lack of standardized orthography, frequent code-switching with French, and a scarcity of annotated data. The authors propose a modular, end-to-end architecture that bridges this gap by integrating several key components into a single, functional voicebot.

The proposed system, named Dziri Voicebot, follows a layered design that separates audio processing from conversational intelligence. The pipeline begins with a Whisper-based automatic speech recognition (ASR) module, fine-tuned on a newly collected telecommunications-specific dataset. The transcribed text is then processed by a natural language understanding (NLU) module, which uses the DziriBERT model within a Rasa-based framework to handle intent classification and entity extraction. For queries outside the scope of predefined business logic, the system employs a retrieval-augmented generation (RAG) module, which uses a Llama 3.2 model to provide contextually relevant answers based on a domain-specific knowledge base. Finally, the system generates spoken responses using a neural text-to-speech (TTS) engine fine-tuned on a dedicated Algerian dialect corpus.

The authors report strong performance across all components of the pipeline. The ASR module achieved a word error rate of 13.74%, while the NLU module reached 98.4% intent classification accuracy and a 93.9% entity-level F1-score. The RAG module demonstrated effective performance in handling open-domain queries, and the TTS system provided stable, natural-sounding speech synthesis. By creating and releasing these dedicated datasets and models, the authors establish a reproducible baseline for future research into end-to-end conversational AI for under-resourced dialects.