chetan-s20/talksync

TalkSync AI is a desktop application that translates spoken conversation between languages in real time. It is built with Python and a graphical interface library called CustomTkinter. The goal is near-instant translation during live conversations, presentations, and online meetings, so that two people speaking different languages can communicate with very little delay between when one person talks and the other hears the translation. The app captures audio from two sources at the same time: your microphone and your computer's system audio. It listens for speech using a neural model called Silero VAD, then transcribes what it hears using an offline speech recognition engine called Faster-Whisper. For translation, it primarily uses the DeepL API, with a local offline fallback called Argos Translate if the internet connection fails. It then synthesizes the translated text back into speech using Sarvam AI for high-quality Hindi or Piper TTS for English, and plays the audio through your speakers or a virtual audio cable. A key use case is routing translated audio into video meeting software like Zoom, Google Meet, Microsoft Teams, or Discord. By using a virtual audio driver called VB-Cable, the translated speech can be sent directly into those platforms as if it were your microphone input. The app includes echo and feedback suppression to prevent speaker audio from looping back into the capture stream. The app requires an NVIDIA GPU with at least 8 GB of VRAM to run the local speech and voice detection models efficiently. Setup involves installing Python dependencies, configuring API keys for DeepL and Sarvam AI in a environment file, and optionally installing the VB-Cable virtual audio driver. Running the app is as simple as executing a Python script. It also keeps a local database of all translation transcripts and offers post-session AI summaries of the conversation. The README provides clear setup instructions and a visual architecture diagram showing how audio flows through the system. The full README is longer than what was shown.