parakeet-ultra-gguf

parakeet-ultra-gguf

Transcribes recordings in 25 European languages offline, using GGUF copies of Moondream's parakeet-ultra for transcribe.cpp. For people who want local speech-to-text with punctuation and optional token timestamps.

Made by handy-computer 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-ultra-gguf holds GGUF conversions of Moondream's parakeet-ultra speech recognition model, prepared by handy-computer for the transcribe.cpp program. The model turns audio into punctuated, cased text in 25 European languages, with optional timestamps for each token. It can stand in for its base model, parakeet-tdt-0.6b-v3, with the same size, languages and interface.

Who it suits: People who want offline transcription on their own computer, and developers who build local speech tools around transcribe.cpp. The files carry the CC-BY-4.0 license inherited from the base model.

Industry | Who uses it | How they use it | Example
Media | Podcasters and editors | Transcribe interviews and episodes without a cloud service | Turning an interview recording into a cased transcript
Software | Developers | Add local speech-to-text to a tool through transcribe.cpp | Calling the command-line program on saved audio
Research | Speech researchers | Compare quantized builds against the full-precision one | Checking the error rate of Q4_K_M against F32

Anyone who needs live transcription of a running stream, or translation into another language, can skip it, because the card rules both out.

Where to use it:

• transcribe.cpp, built from source on Windows, macOS or Linux
• Offline, on your own machine, since the model runs locally
• Windows, macOS or Linux with Python, for downloading the files with Hugging Face's steps

Use case | Who | What they get back
File transcription | Anyone with audio | Punctuated, cased text
Timed transcripts | Editors and developers | Token-level timestamps when requested
Long recordings | Researchers | Text cut at pauses into pieces no longer than half a minute

How to set it up:

• Git, CMake and a C++ build setup, because transcribe.cpp is built from source
• Python and the huggingface_hub package, to download the GGUF files
• ffmpeg, if your audio is not already 16 kHz mono WAV

1. Clone the transcribe.cpp repository from its GitHub page.
2. Build it with the two commands in the first block below, run inside that folder.
3. Install the huggingface_hub package and download the GGUF files with the commands in the second block.
4. Pick one quantization file, such as the Q8_0 one.

cd transcribe.cpp
cmake -B build && cmake --build build

pip install -U huggingface_hub
hf download handy-computer/parakeet-ultra-gguf

How to use it: You run the transcribe-cli program with a model file and a 16 kHz mono WAV. If your audio is in another format, convert it first with ffmpeg.

build/bin/transcribe-cli \
  -m parakeet-ultra-Q8_0.gguf \
  input.wav

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

The card lists six quantizations, with the error rate measured on the LibriSpeech test-clean split.

Title: Quantization options
Type | Size | Word error rate
F32 | 2.51 GB | 1.80%
F16 | 1.26 GB | 1.80%
Q8_0 | 740 MB | 1.80%
Q6_K | 611 MB | 1.81%
Q5_K_M | 550 MB | 1.81%
Q4_K_M | 486 MB | 1.82%

Example: You convert a recording to 16 kHz mono WAV with ffmpeg, then run transcribe-cli with the Q8_0 file on it. You get the spoken words back as punctuated text with normal capitalization.

For newcomers and experts: The first small thing to try is building transcribe.cpp and transcribing one short WAV with the Q8_0 file, which sits in the middle of the size range. It shows the whole path from audio to text.

Inside, a FastConformer encoder feeds a TDT transducer decoder, and a compact speech-detection head marks the pauses where long audio is cut. Decoding is greedy with no external language model. The card says its C++ F32 build matches the reference runtime's score and that each quantization stays close to it.

Convert audio to 16 kHz mono WAV
Run transcribe-cli with a GGUF file
The voice head splits long audio
The model decodes each segment
Get punctuated text

Title: FLEURS word error over 25 languages, in percent
parakeet-tdt-0.6b-v3: 11.62%
parakeet-ultra: 9.55%

Good to know:

• License: CC-BY-4.0, inherited from the base model, with the license text and a notice file in the repository
• Cost to run: none named
• Data it reaches: the audio files you give it, and the download from Hugging Face
• Not for: live streaming or translation, as the card states
• The FLEURS figures come from the original model card, which the card calls the authoritative source

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transcribe.cpp
• Parameters: 627M
• Weights: about 314 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 1,375,499
• Likes: 8
• License: cc-by-4.0
• Languages: en, de, fr, es, it, pt, ru, uk
• Based on: moondream/parakeet-ultra
• Tags: transcribe.cpp, gguf, speech-to-text, parakeet, conformer, multilingual, long-form, automatic-speech-recognition

Links

• Model: https://huggingface.co/handy-computer/parakeet-ultra-gguf
• Files: https://huggingface.co/handy-computer/parakeet-ultra-gguf/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download handy-computer/parakeet-ultra-gguf

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