NVIDIA has launched a 600 million parameter automatic speech recognition model. It is completely open source under CC BY 4.0 and you can use it for commercial work. It handles punctuation, capitalization, and accurate timestamps, so turning audio into subtitles is easy.
Words that usually trip models on my voice, like Cohere and Mistral, were correctly identified. Numbers were handled well too, for example 22 billion parameter model came out with the right numeric form. Across the board, it does a pretty good job.
NVIDIA Parakeet TDT ASR overview
The model is called Parakeet TDT. TDT refers to the architecture, built on Fast Conformer TDT. It is optimized for NVIDIA GPUs and I still need to figure out how to run it on a Mac.
This is strictly an automatic speech recognition model, speech to text. It can detect words and symbols, and it supports very accurate timestamp prediction. It is designed for high quality English transcription with punctuation and capitalization.
NVIDIA Parakeet TDT ASR performance
NVIDIA Parakeet TDT sits at the top of the Hugging Face Open ASR leaderboard. The leaderboard is sorted by WER, the word error rate that measures mistakes. Lower WER is better, and this model ranks above larger instruction and Canary variants in current results.
Whisper, which I use a lot because it is consistent for me, is not in the top five right now. I still use Whisper often, but Parakeet looks like a strong option if you want something that captures different accents well. If you work with Whisper in Python, see this practical Whisper project.
Testing NVIDIA Parakeet TDT ASR
I ran an American accent audio clip through the demo. The transcript preserved capitalization and punctuation and produced accurate timestamps. It handled a sentence that began with “In 1964 I was a little girl,” and it produced proper names like Sidney Poitier correctly.
Use the NVIDIA Parakeet TDT ASR demo
Step 1: Open the official demo space. You can try it here: NVIDIA Parakeet TDT ASR on Hugging Face.
Step 2: Click Upload and select your audio file. Supported accents and pronunciations were handled well in my tests.
Step 3: Wait for the file to finish uploading. The model will start transcribing and show progress in the interface.
Step 4: Review the output with punctuation, capitalization, and timestamps. You can copy the text, create subtitles, or prepare a draft for publishing.
If you want a broader walkthrough on Hugging Face transcription, check this guide to converting English speech to text on Hugging Face. It pairs well with Parakeet’s workflow.
NVIDIA Parakeet TDT ASR for long audio and publishing
It can transcribe long hours of audio. You can take a recording and convert it into a blog post or subtitles with proper punctuation and casing. If you also need speech from your finalized text, explore these open source text to speech options.
NVIDIA Parakeet TDT ASR model details
This is a 600 million parameter speech to text model focused on high quality English transcription. It detects words and symbols accurately and produces strong timestamp predictions. The model architecture is Fast Conformer TDT and it is tuned for NVIDIA GPUs.
Resources
Try NVIDIA Parakeet TDT ASR here: official Hugging Face demo.
Final thoughts
For English transcription with punctuation, capitalization, and timestamps, NVIDIA Parakeet TDT ASR delivers strong accuracy. It handled brand names, numbers, and accents well in my tests. If you need a high quality open source ASR model that is ready for commercial use, this is worth adopting.