Moondream Parakeet Redux

Moondream Parakeet Redux

A 178 MB speech recognizer from moondream that runs at 113 times real time on eight x86 CPU cores. It transcribes English, German, French, Spanish, Italian, Portuguese, Russian and Ukrainian.

Made by moondream on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Parakeet Redux is a speech-to-text model from moondream with 149M parameters. It rebuilds the parakeet-tdt-0.6b-v3 recognizer with the same architecture and tokenizer, but each encoder weight takes one of only three values: minus one, zero or plus one. That ternary design gives 1.58 bits per weight and a 178 MB download.

Who it suits: Developers who want fast speech transcription on an ordinary CPU or laptop, and researchers who study low-bit speech models.

Industry | Who uses it | How they use it | Example
Software | Developers | Add speech-to-text to an app that runs on a CPU | A transcription feature with no GPU
Media | Editors | Turn long recordings into text with timestamps | Captions for a long talk
Research | ML researchers | Study 1.58-bit weights in a speech model | A comparison against the original model
Education | Students | Transcribe lectures in several languages | Notes from a Spanish lecture

If you need the lowest error rate on noisy recordings, you can skip it, since the card shows it behind the original under background noise.

Where to use it:

• Photon, moondream's runtime, which reads the packed weights directly on x86 CPUs, ARM CPUs and Apple GPUs
• Python programs, through the moondream package
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Plain transcription | Developers | The spoken text from an audio file
Sentence timing | Editors | The start and end of each sentence in seconds
Word timing | Subtitle makers | The start and end of every word

How to set it up:

• Python, to run the Hugging Face download tool
• The moondream package, version 2.4.0 or later
• A CPU, an Apple GPU or an NVIDIA GPU, chosen with the device setting

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Install the moondream package, version 2.4.0 or later.
4. Open the model with the photon call and pick a device.

pip install -U huggingface_hub
hf download moondream/parakeet-redux

How to use it: Open the model through the photon call, choose cpu, mps or cuda as the device, and pass an audio file to transcribe. Ask for segment timestamps to get one entry per sentence, or word timestamps to get every word.

import moondream as md

with md.photon("moondream/parakeet-redux", device="cpu") as speech:
    result = speech.transcribe(audio="speech.wav")
    print(result["text"])

Example: You open the model on the cpu and transcribe a file named speech.wav. The result holds the text, and with segment timestamps each sentence arrives with its start and end in seconds.

For newcomers and experts: Start on the cpu with a short recording and plain text output, then turn on timestamps once that works. Expect more errors with background noise, where the card shows it behind the original.

The card compares it with the original on several sets, scored with the Open ASR Leaderboard's own pipeline. Lower word error rate is better. It wins on FLEURS at 10.56 against 11.62 and on long-form TED-LIUM at 2.51 against 2.71. Business speech is 6.96 against 6.15, and noise is 9.04 against 6.72. Its sibling Parakeet Ultra is the full-precision version for GPUs.

Install the moondream package
Open the model with photon
Pass an audio file to transcribe
Text and timestamps come back

Good to know:

• License: CC BY 4.0
• Cost to run: none named
• Data it reaches: downloads files from Hugging Face, then runs on your own machine
• Not for: noisy recordings where accuracy matters most, since the card shows more errors than the original there
• Limits: the card says every speed number it gives comes from Photon, so other runtimes may differ

Title: Real-time speed on eight x86 CPU cores
This model in Photon: 113 times
parakeet.cpp: 45 times
sherpa-onnx: 42 times
onnx-asr: 28 times

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Parameters: 149M
• Weights: about 149 MB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 9,577
• Likes: 320
• License: cc-by-4.0
• Languages: en, de, fr, es, it, pt, ru, uk
• Quantized versions: 12
• Fine-tunes: 7
• Tags: safetensors, parakeet_tdt, ternary, parakeet, speech-recognition, apple-silicon, automatic-speech-recognition, 8-bit

Links

• Model: https://huggingface.co/moondream/parakeet-redux
• Files: https://huggingface.co/moondream/parakeet-redux/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download moondream/parakeet-redux

Then follow the usage in its model card. Nothing in this kit runs by itself.