ClearVoice 2.0

Over 70 million individuals worldwide live with speech dysfluencies such as stuttering, blocks, repetitions, and mid-sentence restarts. In our voice-first digital world, conventional AI systems fail them: traditional Voice Activity Detection (VAD) abruptly terminates recording during brief pauses, speech-to-text creates garbled repetition bloat, and unconstrained language models introduce dangerous factual hallucinations like altering numbers or stripping negations. ClearVoice 2.0 solves this by introducing a real-time, sub-second speech repair and neural resynthesis pipeline. Built with a high-performance polyglot architecture, our Python core engine uses PyTorch Silero VAD with a custom 1,500ms adaptive silence window and 512-sample frame slicing to eliminate premature cutoffs, streaming live audio directly to AssemblyAI's Realtime WebSocket API. When speech is transcribed, our Node.js intelligence server leverages Groq-hosted open-weights models alongside a 6-utterance rolling context buffer to intelligently remove dysfluencies while preserving intentional emphasis. To guarantee zero hallucination, a deterministic Meaning Validator rigorously inspects every repair, ensuring critical numbers, dates, negations, and named entities remain 100% faithful to the original speech. If any drift is detected, an automated zero-trust fallback engine serves a conservatively cleaned or raw transcript. Approved speech is resynthesized through a resilient 3-tier failover engine: primary high-fidelity streaming via ElevenLabs PCM16, secondary on-server system TTS via pyttsx3, and tertiary browser-level client synthesis. Deployed in a Dockerized AWS EC2 backend with an interactive Next.js 16 frontend on Vercel, ClearVoice 2.0 democratizes voice interfaces and empowers dysfluent speakers with clarity, dignity, and confidence.