LingualDub
Composable, registry-based speech-AI framework for low-resource languages
An open, modular speech-AI framework designed as reusable research and production infrastructure for underrepresented languages. Standardizes language metadata, audio datasets, model adapters, and evaluation metrics, with Luganda (lug) as initial validation baseline.
The Challenge
Speech AI pipelines (ASR, TTS, Voice Conversion, Dubbing) have predominantly been built and optimized for high-resource languages like English and Mandarin. Developing speech tools for African languages often meant rebuilding fragile glue-code from scratch for each language, with zero standardized metadata, inconsistent phoneme sets, and tightly coupled model runtimes.
Architecture & Solution
Designed LingualDub from first principles around a decoupled, registry-based architecture. Language metadata, audio datasets, model adapters, execution pipelines, and evaluation metrics are isolated behind standard interfaces. Any researcher or engineer can plug in a new low-resource language adapter or benchmark model without modifying framework internals.
┌────────────────────────────────────────────────────────┐
│ LingualDub Core │
├───────────────────┬───────────────────┬────────────────┤
│ Language Registry │ Pipeline Engine │ Model Adapters │
│ • Luganda (lug) │ • Preprocessing │ • Whisper │
│ • Dialect maps │ • Alignment │ • VITS / TTS │
│ • Phoneme tables │ • Evaluation │ • Custom ONNX │
└───────────────────┴───────────────────┴────────────────┘
│
Automated CI & Validation Suite
(Mypy Strict · Ruff · PyTest)Engineering Decisions
- Registry-Based ExtensibilityImplemented dependency inversion where language packs and models register themselves dynamically, allowing zero-friction community contributions for new African dialects.
- Strict Static Typing & Ruff LintingEnforced 100% strict Mypy type-checking and automated Ruff lints to guarantee industrial reliability across research experimentation.
- Luganda as Initial Ground TruthSelected Luganda (Bantu / Great Lakes) as the reference validation language due to its agglutinative morphology and complex tonal characteristics, proving the framework can handle demanding linguistic structures.
Outcomes
- Live interactive documentation and portal shipped at lingualdub.pages.dev
- Automated CI/CD validation running continuous acoustic and lexical checks
- Standardized interface adopted for ongoing Bantu low-resource speech benchmarks