cth9191

Whisperflow Local — AI skill for Claude Code

AI community

A 100% local WhisperFlow clone for Windows — Ctrl+Win to dictate, faster-whisper (CUDA) + a local Ollama cleanup LLM, focus-safe paste.

How to install Whisperflow Local

This entry records only its repository, not the path inside it, so there is no exact command to give. Open cth9191/whisperflow-local and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Whisperflow Local does

A 100% local WhisperFlow clone for Windows — Ctrl+Win to dictate, faster-whisper (CUDA) + a local Ollama cleanup LLM, focus-safe paste. No cloud at dictation time.

Alternatives in AI

Humanizer-zh: 去除 AI 写作痕迹 — 你是一位文字编辑，专门识别和去除 AI 生成文本的痕迹，使文字听起来更自然、更有人味 5k ★
Azure OpenAI .net — GPT-4, embeddings, DALL-E, and Whisper client 1.8k ★
AI Avatar System — 🎭 AI Avatar / digital human platform — upload a photo, clone a voice, talk to any face in real time with lip 422 ★

README

whisperflow-local

A 100% local WhisperFlow clone for Windows — global hotkey dictation with a **local** AI cleanup pass. Tap a hotkey, speak, and clean corrected text (filler removed, grammar/punctuation fixed) is inserted into whatever text field has focus. STT and the cleanup LLM both run on-device — **no cloud calls at dictation time.**

![whisperflow-local](assets/architecture.png)

Built as the worked example for a YouTube video whose real deliverable is the *process*: deep-research → grill → cross-model plan review → Claude Code builds. See `PLAN.md` (the locked spec), `PLAN-REVIEW-LOG.md` (the Claude↔Codex argument that hardened it), and `docs/architecture.html` (a visual walkthrough).

Pipeline

`Ctrl+Win` (toggle) → mic capture (`sounddevice`, 16 kHz mono) → STT (`faster-whisper large-v3-turbo`, CUDA) → cleanup (local Ollama LLM) → focus-safe clipboard paste into the active field. A non-activating, click-through PySide6 pill shows live mic level without stealing focus; a tray icon gives a clean quit.

Requirements

Windows 10/11
An NVIDIA GPU with a recent driver (the STT model runs on CUDA; CPU fallback exists but is slow)
Python 3.11+ and `uv` (the package manager)
Ollama for the local cleanup model

Setup with Claude Code (recommended)

The fastest path — let Claude Code do the install for you:

git clone https://github.com/cth9191/whisperflow-local.git
cd whisperflow-local
claude            # then ask it to set the project up

Then tell Claude Code:

Set up whisperflow-local on my machine: install the Python deps with uv, make sure Ollama is installed and pull the cleanup model from config.yaml, then run `python -m whisperflow_local doctor` and fix anything that fails until it passes.

The repo's `CLAUDE.md` tells Claude Code exactly how to do this, including the common CUDA/cuDNN gotchas on Windows.