Dziri Voicebot: An End-to-End Low-Resource Speech-to-Speech Conversational System for Algerian Dialect
Dihia Lanasri, Fairouz Taki, Asma Kemmoum
Read on arXiv →
Key claim
A complete speech-to-speech system for Algerian Dialect.
In plain English
Imagine trying to build a conversational system that can understand and respond in a specific dialect, like Algerian Dialect. This dialect poses unique challenges: there’s no standardized way to write it, people often switch between Arabic and French, and there aren’t many resources available to train models. As a result, existing speech technologies tend to overlook these dialects, making them less useful for speakers who rely on them.
Currently, most speech systems are designed for high-resource languages, which means they often fail when faced with dialects that lack sufficient training data. This leads to poor recognition accuracy, misunderstandings, and ultimately, a frustrating user experience. This is what's called a resource bias problem, where the technology doesn't adapt well to the needs of diverse language speakers.
To tackle these issues, the authors developed a modular conversational system specifically for Algerian Dialect. They created dedicated datasets for automatic speech recognition (ASR), natural language understanding (NLU), and text-to-speech (TTS) in the telecom domain. By fine-tuning existing models and integrating them into a single architecture, they aimed to improve the overall interaction quality. The ASR system uses a Whisper-based adaptation, while the NLU combines transformer embeddings with a dialogue framework, and a neural TTS system generates spoken responses from a newly collected dialectal corpus.
The results show that this system performs well across all components, achieving low word error rates for ASR and high scores for intent classification and entity recognition in NLU. This work not only extends previous efforts in dialectal conversational systems but also provides a solid foundation for future developments in speech technologies for low-resource languages.
This work extends existing speech technologies to a low-resource dialect, addressing unique challenges.
The paper provides experimental results across multiple components, demonstrating solid performance.
Deep reliability assessment
The paper supports a component-level feasibility baseline for Algerian Dialect voice bots in a narrow telecom setting, especially showing that Whisper adaptation can reach 13.74% WER on their collected corpus. It overclaims reproducibility and deployability unless the datasets, code, end-to-end task success, latency, robustness to real callers, and downstream error propagation are evaluated or released.
Reproducibility
No public code repository or dataset URL is mentioned in the provided paper text. The authors describe newly constructed ASR, NLU, and TTS datasets, but it is unclear whether they are publicly released, so reproducibility appears limited despite the paper calling the system a reproducible baseline.
Key figure
The key architecture diagram depicts a modular voice-bot pipeline where spoken Algerian Dialect input is transcribed by ASR, interpreted by NLU/dialogue management with RAG support, and converted back into spoken output through TTS.