Willow Uses CoreWeave Infrastructure to Train Dictation Models

Willow Voice is using CoreWeave’s infrastructure to refine the speech-recognition and language models behind its dictation application, with the setup allowing the startup to focus on model evaluation and alignment.

Lawrence Liu, Willow Voice’s co-founder and chief technology officer, said the company uses CoreWeave’s managed reinforcement-learning infrastructure for its training work. Reinforcement learning uses feedback to improve how a model behaves.

“The great thing about CoreWeave’s [reinforcement learning] infrastructure is we don’t have to worry about anything under the hood,” Liu said. “The only thing we really need to focus on at this point is eval alignment.”

Willow fine-tunes its speech model with quiet, real-world audio, including whispered speech and recordings made in coffee shops. Its application is available on Mac, Windows and iPhone.

The system also uses a small language model to process dictated speech after transcription. That model formats and rewrites text produced by the speech-recognition system.

“I think the main part of our secret sauce is we have this really, really small LLM that post-processes everything that you say,” Liu said.

CoreWeave launched its fully managed serverless reinforcement-learning capability in October 2025. The service is designed to reduce the infrastructure and operational work required for reinforcement-learning workloads.

Willow and CoreWeave have not announced financial terms, a contract value or a formal commercial partnership. Liu appeared at CoreWeave’s Fully Connected 2026 event on Sept. 30.

“Within two years, we want it such that you don’t even need keyboards anymore,” Liu said.