Maya Language Tutor

Most language apps teach you to read and recognize a language — not to speak it. Flashcards and grammar drills never touch the actual skill: producing sentences out loud, in real time, under mild social pressure. Human tutors help, but they're expensive and scheduled. Practicing alone gives you silence instead of feedback. Maya fixes that by being a real-time, voice-first conversation partner. Open the browser and start talking — voice activity detection knows when you start and stop, so it feels like an actual phone call. Maya transcribes what you say, replies out loud with an LLM, and the moment you finish speaking, a feedback panel shows three things: What you said (your raw transcript, filler words and all), What you meant (a cleaned-up version of your intended sentence), and More natural (how a native speaker would actually phrase it, with the key change highlighted). You can also select any word mid-conversation for an instant dictionary lookup — definition, part of speech, US/UK pronunciation audio, and example sentences with Chinese translations — and save anything into a personal library to review later. Under the hood: a FastAPI backend, faster-whisper/whisperx for speech-to-text, Fireworks AI or a local Ollama model for reasoning, and Edge TTS or a local CSM-1B model for speech synthesis. Every stage in the loop has a fully local, self-hostable option, so the whole conversation can run offline on-device without a per-token cloud bill. Cached filler audio ("Sure.", "Got it.") plays instantly in Maya's own voice to mask model latency while the real reply generates. Currently supports English and Chinese practice. Built on top of teammate KaushikSiva's original real-voice repo for the 2026 AMD Hackathon.