dilo

An offline-first AI voice assistant that never leaves the browser

RoleIndependent / Personal Project

TypePersonal

Timeline2025

Team1 person

StatusCompleted

RepoPublic

What I did

Built a PWA running Whisper.cpp entirely client-side via WebAssembly

Implemented Web Workers to offload transcription and keep the UI responsive

Built graceful degradation to a Gemini API fallback when offline processing isn't feasible

Implemented Coqui-style TTS synthesis completing the full speech-in → text → speech-out loop client-side

The problem

Voice assistants typically require a server round-trip for every interaction, introducing latency, privacy concerns, and dependency on network connectivity.

The solution

Built a Progressive Web App that runs Whisper.cpp via WebAssembly for speech-to-text directly in the browser. Web Workers offload transcription to background threads. Graceful degradation to Gemini API fallback. Coqui-style TTS handles speech output.

Key features

Whisper.cpp running client-side via WebAssembly
Web Worker-based transcription for responsive UI
Graceful degradation to Gemini API fallback
Coqui-style TTS synthesis (speech output)
Progressive Web App (offline-capable)

Engineering highlights

★WebAssembly + Web Worker Architecture

Running Whisper.cpp via WebAssembly in the browser requires careful memory management and thread isolation. The Web Worker architecture keeps transcription completely off the main thread.

Engineering decisions

Whisper.cpp inference too heavy for the main thread

Offloaded all WASM transcription to dedicated Web Workers with message-passing

UI remains fully responsive during transcription.

Technology stack

Next.jsWhisper.cppWebAssembly

Impact

→Demonstrated low-level systems work (WASM, worker threads) in a full-stack portfolio

Ready to build something exceptional?

Let's discuss how custom software can accelerate your business outcomes.

Start a Conversation