LingoHUD

LingoHUD uses AI models to provide real-time translation and transcription features. These include: (1) On-device speech recognition using OpenAI's Whisper model (runs locally) (2) Offline neural translation using Helsinki University's OPUS-MT models (3) Optical character recognition using RapidOCR and MangaOCR (local) (4) Speech bubble detection in manga using YOLOv8 (local, ONNX inference) (5) Language identification using Meta's FastText (local) Optionally, users may enable online translation services (Google Translate, DeepL, Groq Cloud, Google Gemini) by providing their own API keys. When enabled, data is sent only to the respective service provider; the developer does not collect or transmit user data. All AI processing is triggered by explicit user action (hotkey press). No content is generated in the background. AI outputs (translations, transcriptions, detected text) are probabilistic and may contain errors. The software is intended for casual assistive use in gaming, manga reading, and language learning, and is not suitable for legal, medical, or mission-critical applications. Built-in safeguards include hallucination detection to filter common Whisper errors, smart filters to skip non-translatable content (URLs, FPS counters, timestamps), and noise gating to prevent false transcriptions from silence.