Parakeet is a family of automatic speech recognition (ASR) models, models that convert spoken audio into written text, published by NVIDIA as open model weights on Hugging Face. parakeet-tdt-0.6b-v2 transcribes English; parakeet-tdt-0.6b-v3 extends it to 25 European languages and detects the language of the audio automatically, without a prompt. Each model has 600 million parameters.
The architecture combines a FastConformer encoder, a neural network that turns the audio signal into a compact sequence of features, with a TDT decoder, a transducer that predicts each text token together with its duration. The input is 16 kHz audio in .wav or .flac format. The output is text with punctuation and capitalization, plus word-level and segment-level timestamps. On an A100 80GB GPU, one pass can transcribe up to 24 minutes of audio with full attention, or up to 3 hours with local attention. The models run with NVIDIA NeMo, with Hugging Face Transformers, or with NeMo-Speech.cpp, a lightweight C++ runtime for local inference.
The weights are free to download under the CC BY 4.0 license, and the model card says they are ready for commercial and non-commercial use.