parakeet-tdt-0.6b-v3

parakeet-tdt-0.6b-v3

Transcribes speech in 25 European languages, detects the language by itself and adds punctuation, capitalization and word timestamps. A 600M-parameter speech-to-text model from NVIDIA for developers and transcription services.

Made by nvidia 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: NVIDIA's parakeet-tdt-0.6b-v3 turns speech into text and has roughly 600 million parameters. It extends the earlier English-only v2 model to 25 European languages. It finds the language of the audio on its own and transcribes it without extra prompting. It adds punctuation and capital letters and returns timestamps for words and segments.

Who it suits: Developers and teams who build transcription, voice assistants, subtitles or voice analytics and need one model for many European languages.

Industry | Who uses it | How they use it | Example
Media | Subtitle teams | Generate captions and subtitles with timestamps | Subtitles for a Czech interview
Customer service | Voice analytics teams | Transcribe calls for later analysis | Text from a French support call
Software development | App developers | Add speech input to an app | A voice assistant in Spanish
Education and research | Academics | Transcribe lectures and recordings | A German lecture turned into text

If you only need languages outside its 25, you can skip it.

Where to use it:

• Python with the transformers library, on Windows, macOS or Linux
• Transcription services and voice assistants that you build
• Subtitle and analytics pipelines

Use case | Who | What they get back
Transcription | Service builders | Punctuated text with word and segment timestamps
Subtitles | Video teams | Timed text that you can turn into captions
Long recordings | Researchers | Transcripts of long audio, with a higher limit under local attention

How to set it up:

• Python and pip
• About 2.5 GB of disk space in F32, or about 314 MB at 4-bit (an estimate)
• Audio files in a language it supports

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Load the model with transformers or another supported tool, following the model card.

pip install -U huggingface_hub
hf download nvidia/parakeet-tdt-0.6b-v3

How to use it: Give the model an audio file and read back the transcript. It works out the language by itself, so you do not name it. Punctuation, capitalization and timestamps come with the text.

Example: You pass in a recorded meeting in French. The model identifies the language on its own and returns the text with punctuation and capital letters, plus the time each word and segment begins.

For newcomers and experts: If you are new to speech recognition, try one short recording in a language you know and compare the text with what you hear. Start with clear audio.

The model belongs to the FastConformer-TDT family and was trained mostly on the Granary multilingual corpus. The card says Portuguese scores can look off, since training used European Portuguese and most test sets use the Brazilian kind. Long audio is handled with full attention on an A100 80GB or with local attention.

You provide an audio file
The model detects the language
It transcribes with punctuation
You get text and timestamps

Good to know:

• Language scope: 25 European languages, including English, Spanish, French, German and Russian
• Portuguese: results may differ from benchmarks because of European and Brazilian variants

Title: Longest audio per recording by attention type
Full attention on an A100 80GB: 24 minutes
Local attention: 3 hours

• License: cc-by-4.0, with commercial and non-commercial use allowed
• Cost to run: no API key or paid service is named, only your own hardware
• Data it reaches: downloading fetches model files, and it reads the audio you give it
• Not for: languages outside the 25 it supports

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 627M
• Weights: about 2.5 GB in F32; about 314 MB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 529,594
• Likes: 1,272
• License: cc-by-4.0
• Languages: en, es, fr, de, bg, hr, cs, da
• Quantized versions: 145
• Fine-tunes: 140
• Tags: transformers, nemo, safetensors, gguf, parakeet_tdt, feature-extraction, automatic-speech-recognition, speech, audio, transducer, transformer, fastconformer

Links

• Model: https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
• Files: https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/parakeet-tdt-0.6b-v3

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