LocalWhisper •LocalWhisper •

LocalWhisper came from a personal problem: I wanted to practice spoken English on my way to school, say phrases out loud, and receive corrections, but I did not feel comfortable doing that in public. I also did not want my voice or conversation history sent to external servers. The app had to feel private, run without internet, and keep the whole interaction on my machine. The challenge was connecting three AI models locally on normal consumer hardware: OpenAI Whisper for speech-to-text, Llama 3.2 3B as the LLM for feedback and corrections, and Kokoro TTS for voice output. Keeping everything local was the defining decision. External APIs would have been simpler, but they would have broken the privacy promise. The cost was technical complexity: performance varies a lot between machines, and the local LLM is the slowest part of the pipeline. Built in Python with Flet in 10 days. It works as a complete MVP with all three models connected. The next iteration will likely move the interface to React or Next.js while keeping the local AI engine intact.