DEV Community
Follow
Dicter du code confidentiel : pourquoi j'ai arrêté d'envoyer ma voix dans le cloud
An AI consultant expresses concern about the privacy implications of cloud-based dictation tools. These services send audio data containing sensitive client information to remote servers. This data could be used for model improvement or accessed internally, posing a risk to confidential business details. As the consultant is bound by NDAs, they are solely responsible for any data breaches.To address this, the consultant advocates for 100% local dictation solutions. These solutions process audio entirely on the user's machine, ensuring no data leaves the device. A simple test of disconnecting Wi-Fi can verify if a dictation tool is truly local.The underlying technology for local dictation often uses faster-whisper, an optimized implementation of OpenAI's Whisper model. This model can run on a user's CPU and requires only a one-time download. Advanced users can integrate this model with custom scripts, including a feature to initially prompt the model with technical terms to improve transcription accuracy.The author acknowledges that fully local solutions have limitations. Transcription accuracy for regional accents may be lower, and the user interface might be less polished than commercial offerings. Currently, the implemented solution is Windows-only.Driven by the desire for a user-friendly local dictation experience, the consultant packaged their setup into an installer. This installer simplifies the technical process for users. The software is open-source, allowing for transparency regarding its operation.A paid, one-time purchase version is available, emphasizing a commitment to privacy and avoiding subscription models. The author is currently building this solution in public, seeking user feedback. A waitlist is provided for those interested in the technology.