Nova — Voice AI Agent
A bilingual (Turkish/English) voice assistant that runs fully locally on open-source components: speech recognition, RAG over company documents, a local LLM and text-to-speech.
About this project
I built an end-to-end voice assistant that listens to the user, detects the spoken language (Turkish/English) on its own and answers out loud in the same language. The pipeline combines Silero VAD, Whisper speech recognition, bge-m3 based RAG, a local Qwen2.5 language model and Piper text-to-speech, all running locally. It takes facts from sources rather than its own memory: company documents, a live SQL product database, and weather and market APIs. Code-based calculation and number verification layers block hallucinated facts. It scored 27/27 on the accuracy and hallucination evaluation, 38/38 on multi-turn conversation scenarios and 24/24 on language detection, with a median perceived latency of 2.8–4.2 s.
The Problem
Off-the-shelf voice assistants depend on cloud services, are hard to ground reliably in company data and often hallucinate facts.
The Solution
I built a fully local, open-source voice assistant that grounds its answers in documents, a SQL database and APIs, and verifies numbers with code.
Key Features
Screenshots
