The author developed a local transcription tool for macOS to process meeting recordings without relying on cloud services. The project involved testing 32 configurations of various speech-to-text models, including Whisper, Parakeet, and GigaAM, to ensure data privacy for sensitive business discussions.
The evaluation process utilized a dataset of 36 hours of meeting recordings and a smaller benchmark set of 82 fragments. The author measured performance using the Word Error Rate (WER) metric, focusing on the accuracy of Russian speech and technical terminology.
Initial testing showed that GigaAM performed best for general Russian speech, while Parakeet was superior at identifying English technical terms. The author implemented a custom phonetic metric, WERx, to distinguish between spelling errors and actual word misinterpretations.
The final solution uses a cascade of two models: GigaAM provides the base transcription, while Parakeet is used to correct technical terms. This approach improved the WER from 27.0% to 21.7% on the author's test datasets and increased the accuracy of technical term transcription from 19% to 44%.
