Inza01/ivoirevoice-ai

French–Dioula automatic speech recognition platform with Ivorian accent adaptation, low-resource ASR, LoRA fine-tuning and controlled post-correction. # IvoireVoice AI Local, reproducible automatic speech recognition experiments for French and Dioula in the Ivorian context. IvoireVoice AI provides a Gradio comparison dashboard, a speaker-disjoint data pipeline, Whisper baselines, a full Dioula Whisper Tiny adaptation, ASR metrics and privacy-aware exports. The project is a research MVP, not an industrial-ready transcription service. An isolated Phase 2 web platform lives under `web/` and coexists with the local Gradio demo. Its transcription page is connected to a versioned local FastAPI adapter and the existing ASR service. Text translation now has a local-only NLLB research adapter and remains explicitly `experimental`; learning content, authentication and persistence remain `coming_soon`. > **Résumé en français —** IvoireVoice AI est un démonstrateur local de > transcription français–dioula. Il compare Whisper Tiny, Whisper Small et un > Whisper Tiny adapté sur des données dioula locales. Le corpus réel, les > transcriptions individuelles et le checkpoint ne sont pas publiés. ## Problem General-purpose ASR models have limited coverage of low-resource West African languages and local acoustic conditions. This project explores whether a small, reproducible Whisper adaptation can improve Dioula transcription while keeping: - speaker identities and audio private; - train, validation and test speakers strictly separated; - experimental claims limited to the evaluated data; - model loading and execution possible on a local machine. ## Objectives - inventory and validate a local Dioula speech corpus; - build deterministic speaker-disjoint splits; - compare public Whisper Tiny and Whisper Small baselines; - adapt Whisper Tiny through a bounded pilot, full development and frozen refit; - compare baseline and adapted outputs in a local interface; - report WER, CER, latency, RTF and edit counts; - keep the architecture extensible to Baoulé and other Ivorian languages. ## MVP features - upload WAV, MP3, FLAC or …