parakeet-ultra

parakeet-ultra

Transcribes speech to text with word and sentence timestamps. A post-trained version of NVIDIA parakeet-tdt-0.6b-v3 from moondream, built for developers who need accurate transcripts in many languages.

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-ultra is a speech recognition model from moondream. It turns audio into text and can add start and end times for each sentence or word.

It is a post-trained version of NVIDIA parakeet-tdt-0.6b-v3 with the same architecture, tokenizer and roughly 0.6B parameters. The authors report lower error rates than the original on their benchmarks.

Who it suits: Developers who add transcription to an app, and teams that work with recorded meetings, calls, podcasts or talks.

Industry | Who uses it | How they use it | Example
Software | Developers building voice features | Call the model from Python on audio files | Turn an uploaded recording into text
Media | Podcast and video teams | Transcribe episodes with word timings | Subtitles aligned to each spoken word
Business | Teams with meeting or call recordings | Convert audio into searchable text | A written record of a recorded meeting
Education | Lecture and talk archives | Transcribe long recordings | Text versions of recorded talks

Skip it if you want a finished transcription app, because the source shows only a Python route.

Where to use it:

• Python scripts that use the moondream package and its Photon runtime
• NVIDIA GPU machines, where every accuracy figure on the card was measured
• Windows, macOS or Linux with Python, to download the model files

Use case | Who | What they get back
Plain transcription | Developers | The spoken text of an audio file
Sentence timestamps | Subtitle makers | One entry per sentence with start and end in seconds
Word timestamps | Editors and indexers | Start and end times for every word

How to set it up:

• Python and pip
• An NVIDIA GPU if you want the setup the authors measured
• The moondream Python package, which provides the Photon runtime

1. Install the moondream package.
2. Optionally download the model files with the Hugging Face tools.
3. Put an audio file such as speech.wav next to your script.

pip install moondream

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

How to use it: Open the model with md.photon, then call transcribe on an audio file. Setting timestamps to segment gives one entry per sentence, and word adds the timing of every word.

# pip install moondream
import moondream as md

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

    # timestamps="segment": one entry per sentence, with its start
    # and end in seconds; "word" adds the start and end of every word
    result = speech.transcribe(audio="speech.wav", timestamps="word")
    for segment in result["segments"]:
        print(segment["start"], segment["end"], segment["text"])
        for word in segment["words"]:
            print(word["start"], word["end"], word["word"])

Tool | What it does
md.photon | Opens the model so you can transcribe audio
speech.transcribe | Transcribes an audio file and returns a result
timestamps option | Set to segment for sentences or word for every word
result text | The full transcript as one string
result segments | Sentences with start and end times, and words when requested

Example: You pass speech.wav to transcribe with timestamps set to word. The result holds the whole transcript first. After it come the sentences with start and end seconds, and each sentence lists its words with their own timings.

For newcomers and experts: If you are new to speech recognition, think of this as software that listens to a recording and types it out. Start by installing the package and transcribing one short wav file, then print the text.

If you know the field, this is a post-trained parakeet-tdt-0.6b-v3 that keeps the original architecture and tokenizer. The card says it runs faster in Photon than the original does in NeMo, with 9,743 times real time on LibriSpeech test-clean against 6,005 for the original on one NVIDIA B200.

Audio file such as speech.wav
Photon loads parakeet-ultra
transcribe returns the text
Optional timestamps per sentence or word

Good to know:

• The accuracy and speed numbers come from the model authors and were measured with Photon on an NVIDIA GPU
• On one set, the cleaned VoxPopuli test, the new model scored 3.10 against 3.02 for the original, so it is not better everywhere
• Weights: about 1.3 GB in F16, according to Hugging Face

Title: Average word error rate in percent, lower is better
Original on 7 English sets: 6.26 percent
Ultra on 7 English sets: 5.80 percent
Original on FLEURS, 25 languages: 11.62 percent
Ultra on FLEURS, 25 languages: 9.55 percent

• License: CC BY 4.0
• Cost to run: none named
• Data it reaches: the audio files you pass in, plus the model files you download
• Not for: CPU or Apple silicon use, where the card points to the smaller Parakeet Redux sibling

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Parameters: 627M
• Weights: about 1.3 GB in F16; about 314 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 51,675
• Likes: 142
• License: cc-by-4.0
• Languages: en, de, fr, es, it, pt, ru, uk
• Quantized versions: 53
• Fine-tunes: 26
• Tags: safetensors, parakeet_tdt, parakeet, speech-recognition, automatic-speech-recognition

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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