This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass TouchCall is a voice-first AI assistant designed to help people spend less time staring at screens and more time living in the real world.
Most productivity and wellness apps ask us to open yet another app, check another notification, and spend even more time on our phones. I wanted to try a different approach: what if technology helped us step away from technology?
With TouchCall, users can speak naturally with an AI assistant that encourages them to take a walk, exercise outdoors, or complete a simple real-world activity. Instead of requiring users to keep looking at a screen, TouchCall is built around voice interaction.
The idea is simple: AI should sometimes give us a reason to put our phones down.
I'm building TouchCall as a lightweight, open-source project with a focus on local AI inference and accessible voice interaction.
💻 Project repository: OutVoice on GitHub The current prototype demonstrates the core voice pipeline: speech recording, speech-to-text transcription, local LLM response generation, and text-to-speech playback.
A deployed demo or live phone-call demonstration is not yet linked here.
🔗 GitHub repository: https://github.com/Pranavms09/OutVoice The project brings together open-source speech recognition, local language-model inference, and speech synthesis to create a voice-first assistant.
The goal is to make the system modular so that developers can inspect it, modify it, and experiment with different open-weight models.
I built the initial TouchCall prototype in Python, running locally on my Ubuntu laptop.
The main components are:
llama.cpp.
The current interaction pipeline looks like this:
Microphone → Whisper → Local Qwen model → Piper TTS → Speaker
I also implemented silence-based recording so the system can stop recording shortly after the user finishes speaking instead of always waiting for a fixed recording duration.
One of the interesting engineering challenges was getting the components to work together reliably. I had to address audio recording behavior, model output parsing, and text-to-speech audio generation.
The prototype currently runs on my laptop. My next steps are to improve multi-turn conversation context and integrate scheduled outbound phone calls so users can receive reminders without needing to open an app.
For a project like TouchCall, open innovation is about more than saving API costs. It's about having control over how the assistant works.
Using open-source tools and an open-weight language model lets me run the core AI pipeline locally, experiment with different models, and understand how each component behaves. I'm not locked into a single hosted LLM provider or a proprietary conversational API.
Local inference also creates an opportunity to keep voice transcripts and conversations on the user's machine, depending on how the final system is deployed and configured.
Most importantly, open tools make experimentation possible for individual developers. I can build a voice assistant, inspect its behavior, change its architecture, and share improvements without needing access to a closed AI platform.
Some future features, such as outbound phone calls, may still require an external telephony provider. The goal is to keep the AI core open and replaceable wherever practical.
Open innovation gives developers the freedom to build technology around people, rather than forcing people to adapt to a closed platform.
I'll add any additional partner prize categories if the project meets their specific eligibility requirements.
Built by Pranav Suryawanshi 🌱
The goal of TouchCall is simple: use AI to help people reconnect with the world outside their screens.