Local Audio Transcription: Building an Efficient Model Cascade

The author shares their experience developing a local macOS application for audio transcription, designed to automate meeting minutes. Faced with the need for high-quality speech recognition without relying on cloud services, the developer tested 32 configurations using Whisper, Parakeet, and GigaAM models. The result was an optimal cascade of two models: one generates the base text, while the second corrects inaccuracies. This approach reduced the Word Error Rate (WER) from 27.0% to 21.7%, fixing one in five errors. The article details the model selection process, measurement methodology, and provides practical advice, including a prompt for implementing a similar solution independently. The project is aimed at users who prioritize data privacy and the ability to process audio offline, avoiding the need to send recordings to external servers.