NVIDIA Parakeet

Parakeet TDT 0.6B v2 is NVIDIA's 600M parameter English speech recognition model pairing a FastConformer encoder with a Token-and-Duration Transducer decoder, a combination tuned for throughput: NVIDIA reports an RTFx of 3,380 at batch size 128, and the model topped the Hugging Face Open ASR Leaderboard at 6.05 percent average word error rate on its May 2025 release, with 1.69 percent on LibriSpeech test-clean. Output includes automatic punctuation, capitalization, and accurate word-level timestamps, it copes with spoken numbers and song lyrics, and a single pass can transcribe audio segments up to 24 minutes long. It runs through the NeMo toolkit, installed with pip install nemo_toolkit with the asr extra, and accepts 16 kHz mono audio; NVIDIA validates it on GPUs from V100 and T4 through A100, H100, and L40, though the checkpoint itself needs only about 2 GB of memory to load. The weights are released under CC-BY-4.0, so commercial use is allowed with attribution, and a multilingual v3 covering 25 European languages has since followed.