Nvidia drops Parakeet-TDT-v2: Transcribes an hour of audio in one second

Parakeet-TDT-0.6B-v2 is Nvidia's new open-source automatic speech recognition model. It packs 600 million parameters and does one thing exceptionally well: transcribing an hour of audio in just one second when running on Nvidia GPUs.

The model currently sits at the top of the Hugging Face Open ASR Leaderboard with an average Word Error Rate (WER) of 6.05%. While proprietary powerhouses like OpenAI's GPT-4o-transcribe (2.46% in English) and ElevenLabs Scribe (3.3%) still hold a lead in raw accuracy, Parakeet closes the gap significantly. Best of all, it ships under a commercially permissive CC-BY-4.0 license, meaning anyone can bake it into commercial software for free.

What it can do

Under the hood, Parakeet combines a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. In plain terms, predicting tokens and their durations simultaneously cuts down decoding steps and supercharges inference speed.

Running with a batch size of 128, the model clocks a Real-Time Factor (RTFx) of 3386.02. Beyond raw transcription speed, it provides proper punctuation, capitalization, and precise word-level timestamps right out of the box, making it ready for subtitle generators and meeting summaries.

Robustness looks solid across the board. Tested across benchmarks like AMI, Earnings22, GigaSpeech, and SPGISpeech, it maintains high accuracy even under telephony-style audio and noisy conditions.

Nvidia trained the model on its Granary corpus, which spans roughly 120,000 hours of English audio. That includes 10,000 hours of human-transcribed data and 110,000 hours of pseudo-labeled speech gathered from sources like LibriSpeech, Mozilla Common Voice, and YouTube-Commons. The dataset is slated for public release following Interspeech 2025.

How to run it

Getting started is straightforward. Developers can pull the weights directly from Hugging Face or deploy via Nvidia's NeMo toolkit, which plugs cleanly into standard Python and PyTorch setups.

Hardware-wise, Parakeet is heavily tuned for Nvidia chips like the A100, H100, T4, and V100. While modern data center GPUs unleash maximum throughput, the modest 600M footprint means it can load on systems with as little as 2GB of RAM, making testing accessible.

Why it matters

If you build voice agents, subtitle tools, or transcription pipelines, this release is worth an immediate look. Up until now, teams had to choose between paying hefty per-minute fees for closed APIs or wrestling with sluggish open-weight setups. Parakeet delivers production-grade speed without the billing meter.

Nvidia's play here is obvious: open-source software drives GPU demand. By releasing a top-tier model that runs best on its own silicon, Nvidia gives developers a compelling reason to stick with its hardware ecosystem while putting real pressure on commercial transcription vendors.