This thesis addresses the problem of fully local automatic transcription of clinical interviews without the need for an internet connection. The aim was to experimentally compare modern speech-to-text models, select the most suitable architecture, and subsequently design and implement a functional prototype of the MedApp application, capable of processing real audio recordings of doctor-patient interviews and producing a structured report.

In the first part of the thesis, selected variants of the Whisper model and the Vosk system were tested on fourteen recordings divided into two datasets, using the WER, CER, and RTF metrics. Five recordings consisted of acted clinical recordings, referred to as dataset A, while nine recordings were synthetic recordings created using ElevenLabs, referred to as dataset B. The tests showed that the Whisper large-v3 model achieved the best results, with an average WER of 17.7% and CER of 12.8% without contextual prompting, while also demonstrating low computational time and stable performance when running on a GPU. Based on these results, it was selected as the core of the proposed system.

The second part of the thesis focused on the implementation of a complete on-premise system comprising speech-to-text transcription, speaker diarization using the pyannote.audio 3.1 model, and generation of structured output in JSON format. The solution also includes a module for automatic interview summarization, implemented using the local language model Ollama llama3.1:8b, which creates a structured record containing key topics, a summary, and action points based on the transcript, without the need to send data outside the device.

The proposed system achieved an average WER of 8.2% and CER of 4.9% across all recordings. On the acted clinical recordings in dataset A, it achieved a WER of 15.7%, while on the synthetic recordings in dataset B, it achieved a WER of 3.9%. The average real-time factor reached RTF = 0.460. Speaker diarization using the pyannote.audio 3.1 model achieved an average DER of 2.1%. The outcome of the thesis is a functional and practically validated application prototype that enables fully local transcription of clinical interviews with automatic speaker identification and automatic generation of a structured record.