UIWhisperCPP

A desktop native app for transcribing audio files. Runs locally, no cloud, no API keys.

Screen.Recording.2026-02-26.at.23.32.10.mov

---

UIWhisperCPP wraps local speech-to-text engines - whisper.cpp and Parakeet - in a clean Qt interface - pick your files, hit go, get transcripts. Everything runs on your machine. Your audio never leaves your computer.

What it does

• Transcribe .wav, .mp3 and .m4a files - select one or many, it'll work through them
• Choose your model - Whisper sizes (base to large-v3) or Parakeet v3, a fast multilingual model for Apple Silicon
• Real-time progress - watch the transcription happen with a progress bar and live log output
• Timestamped output - saves transcripts as .txt files with [HH:MM:SS.mmm --> HH:MM:SS.mmm] timestamps (SRT-style)
• Runs completely offline - powered by Whisper or Parakeet, no internet required after the initial model download

Getting started

Requirements

• Python 3.11+
• An Apple Silicon Mac (only required for the Parakeet model; Whisper runs anywhere)
• uv (recommended) or pip

FFmpeg is bundled via imageio-ffmpeg, so no separate install is needed.

Install

git clone https://github.com/olchyk98/UIWhisperCPP.git
cd UIWhisperCPP

# using uv (recommended)
uv sync

# or using pip
pip install -e .

Run

# if installed with uv
uv run app

On first launch, the app will download a Whisper model. This takes a bit depending on your connection - after that, it's cached locally.

How it works

Pretty straightforward pipeline:

1. You pick audio files through the file dialog
2. Audio gets resampled to 16 kHz (what the models expect)
3. The selected backend processes the audio locally - whisper.cpp for Whisper, MLX for Parakeet
4. Segments come back via callbacks - you see them in the log as they arrive
5. Final transcript gets saved as a .txt next to the original file

A small abstraction layer (uiwhispercpp.models) sits behind all of this: each engine implements a common Model interface, and a single ModelManager lists the available models and runs transcription. Adding a model from another package later is just a new Model implementation registered with the manager - nothing else in the app changes.

Built with

• whisper.cpp - fast C++ inference for OpenAI's Whisper model
• pywhispercpp - Python bindings for whisper.cpp
• parakeet-mlx - NVIDIA Parakeet models on Apple Silicon via MLX
• PySide6 - Qt for Python
• imageio-ffmpeg - bundled FFmpeg for audio decoding
• uv - Python package management

License

AGPL-3.0