Multilingual ASR

Problem & context

Speech models are trained overwhelmingly on high-resource languages. For a language like Kinyarwanda, off-the-shelf systems are either unavailable or inaccurate—which locks its speakers out of voice interfaces, transcription tooling, and translated media.

Technical approach

The system pairs speech recognition with translation in a single pipeline, built on PyTorch and Hugging Face Transformers and deployed as a FastAPI service.

• Transformer-based ASR achieving 82–90% transcription accuracy
• GPU-accelerated inference for near-real-time performance
• Redis caching to avoid recomputing repeated segments
• Async processing pipeline sustaining 3× real-time throughput

Decisions & tradeoffs

Accuracy and latency pull against each other in speech systems. The caching and async design let the service keep throughput high without degrading model quality—an optimization problem that mattered more than squeezing out marginal accuracy points, because a system too slow to use helps no one.

Outcome & impact

A working, open-source system that makes English–Kinyarwanda speech accessible at practical speed—an infrastructure piece for underserved linguistic communities, and a template for extending speech tooling to other low-resource languages.

Lessons learned

• Low-resource ML is a systems problem as much as a modeling problem.
• Caching is the cheapest accelerator you will ever deploy.
• Language access is impact you can measure in who gets to participate.

Links

• View on GitHub