GitHub radar
Tiny speech-to-text model that runs on a CPU
Moondream's compressed version of NVIDIA's Parakeet speech model fits in 178 MB, transcribes 25 languages, and runs fast on a regular laptop CPU, no GPU needed.
Parakeet Redux is a lightweight version of NVIDIA's speech recognition model parakeet-tdt-0.6b-v3. It was released by the team Moondream. The architecture and the tokenizer are the same as in the original. But the encoder weights are compressed down to three possible values per parameter, minus one, zero or plus one. That brings the model down to about 178 megabytes instead of the original's 1.2 gigabytes. It recognizes speech in 25 languages, including Russian, and can return text with timestamps per sentence and per word. On eight cores of a regular CPU the authors measured a speed of 113 times real time, without a GPU. On English the model stays close to the accuracy of the original NVIDIA model. On the multilingual test and on long recordings it is even more accurate. In noisy audio the gap to the original widens compared to clean recordings, the authors state this honestly.
Why a vibe-coder should care
If a vibe coder is pulling content out of videos and podcasts, transcribing speech to text is usually the first step. That usually means either a paid cloud service or a heavy model that loads down a GPU. This model is small and fast enough to run transcription locally and in batches, without sending audio to someone else's cloud. For clean speech, podcasts, interviews, lectures, the quality is close to the full-size NVIDIA model. For noisy recordings expect more mistakes and check the output.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the local speech-to-text model Parakeet Redux from Moondream: https://huggingface.co/moondream/parakeet-redux — install it via the moondream Python package, run a transcription on a sample audio file and show me the result with timestamps.
A regular laptop, no GPU needed — the model is about 180 MB and runs fast on CPU alone.
Open on Hugging Face▌ More finds