This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
LocalWhisper Pro is an open-source, offline-first voice interface designed to minimize screen time. Instead of spending hours hunched over a keyboard typing emails, documentation, or field notes at 40 WPM, users dictate naturally at 150+ WPM. Local AI instantly cleans, restructures, and auto-types the text directly at the cursor—enabling users to wrap up screen work fast and get outdoors.
How it gets people off the screen and into the world:
- Makes the screen the shortest interaction: Speech is up to 4x faster than typing. By handling filler-word scrubbing, formatting, and summarization automatically, users spend a fraction of the time staring at word processors and code editors.
- 100% Offline Trail & Field Ready: Because both transcription and refinement run entirely on-device, you can take your laptop outside—onto the porch, into a garden, or down a hiking trail without cell signal—to dictate field logs, nature observations, and ideas without internet dependencies.
-
No Clipboard or Cloud Distractions: It uses low-level hardware stroke injection (
SendInput/pynput) rather than pasting through clipboard history, keeping workflows seamless and distraction-free.
Demo
-
Direct Download: Download the standalone Windows executable (
LocalWhisper_Pro_v3.1_Windows.zip) directly from our GitHub Releases page—no Python runtime required.
Code
Repository: [https://github.com/Harsh-Dev07/Local-WhisperFlow]
License: Open Source (MIT)
How I Built It
LocalWhisper Pro is architected entirely around open-source AI models and local runtime frameworks:
Local Speech Recognition (
faster-whisper/ CTranslate2):
Audio input is captured locally usingsounddeviceand processed throughfaster-whisper, a reimplementation of OpenAI's Whisper model running on the CTranslate2 inference engine. This delivers up to 4x faster-than-real-time transcription on consumer CPUs/GPUs without external network calls.-
Local Intelligence & Text Refinement (Ollama +
llama3.1:8b):
Raw transcription often suffers from stutters, repetitions, and unstructured ramblings. LocalWhisper integrates directly with local Ollama instances. Using open-weight models likellama3.1:8b, it cleans transcripts across multiple modes:- Smart Dictation: Removes filler words while maintaining personal voice.
- Field & Bullet Summary: Distills stream-of-consciousness rambling into structured, actionable checklists.
- Code & Tech: Synthesizes voice logs into camelCase, snake_case, and terminal commands.
- Hinglish/Hindi Processing: Accurately interprets code-switched multi-language dictation.
System Overlay & Direct Keystroke Emulation:
The interface is a lightweight, non-stealing glassmorphic HUD built in Python (Tkinter). Refined text bypasses the operating system's clipboard (Win+V) and is typed directly into active windows via hardware-level event hooks (pynput), ensuring zero data retention outside the user's encrypted local history.
Why Does Open Innovation Matter?
Open innovation was not an afterthought for this build—it was a prerequisite:
-
True Offline Portability: Closed-source voice-to-text platforms (like Google Speech-to-Text or OpenAI Whisper API) fail the moment you walk into a park, mountain trail, or remote campsite without Wi-Fi or cellular service. Open-weight models (
faster-whisper+ Llama 3.1) run autonomously on consumer laptops off the grid. - Complete Sensory Privacy: Voice notes recorded outdoors or in private moments often include personal thoughts, journal entries, and sensitive project ideas. Open-source local models guarantee that no spoken audio or transcribed text ever hits corporate telemetry servers or third-party training pipelines.
- Zero Ongoing Cost: Proprietary APIs charge per audio minute and per token, which discourages long, ambient voice captures. Open-weight inference costs nothing to run, whether you record a two-second note or an hour-long outdoor brainstorm.
-
Model Modularity: Users have the freedom to swap out the refiner engine for smaller models (e.g.,
phi3,gemma2:2b) on low-power devices, or larger parameter models on dedicated hardware.
Prize Categories
- Hacktoberfest Open-Source AI Challenge: Week 1 (Touch Grass)
Top comments (1)
typing through SendInput instead of pasting is a nice call for not clobbering the clipboard. how does it behave in apps that sit behind an IME or drop keystrokes when long refined paragraphs get injected fast, do you throttle the key events at all?