Parakeet-TDT-0.6B-v2

Parakeet-TDT-0.6B-v2, released May 1, 2025, is NVIDIA's open automatic speech recognition model: 600M parameters combining a FastConformer encoder with a TDT decoder, trained on the 120,000-hour Granary dataset. It tops the Hugging Face Open ASR Leaderboard at 6.05% word error rate, transcribes an hour of audio in about one second on NVIDIA GPUs (RTFx 3386), and supports punctuation, capitalization, and word-level timestamps. It is released under CC-BY-4.0, so commercial use is allowed.

The honest picture: it is English-focused (the Granary corpus is English audio), so multilingual transcription still belongs to Whisper. The headline "one second per hour" needs batched A100/H100-class hardware; on modest setups it is fast but not magical. And CC-BY-4.0 requires attribution, unlike Apache/MIT. For English transcription services, voice assistants, and subtitle pipelines where you want open weights without Whisper's heft, it is currently the benchmark to beat.