What it does
Parakeet is a collection of automatic speech recognition models from NVIDIA, published on Hugging Face as NeMo models. The collection description says they reach strong recognition accuracy while staying efficient at inference, and that they come in CTC and RNN-Transducer variants. The collection lists more than a dozen checkpoints, ranging from about 110M to 1.1B parameters. Model names also show TDT variants, a few checkpoints labelled for Japanese and Vietnamese, and streaming or real-time ones, including a multitalker streaming model. A hosted demo for Parakeet-TDT-0.6b-V2 transcribes audio files with timestamps and downloadable subtitles. The collection page does not state licence terms, pricing or supported operating systems, so check each model card before relying on a checkpoint.
Capabilities
- Speech recognition checkpoints from about 110M to 1.1B parameters
- CTC, RNN-Transducer and TDT model variants
- Streaming and real-time checkpoints, including a multitalker one
- Checkpoints labelled for Japanese and Vietnamese
- Demo space that returns timestamps and downloadable subtitles
When to use it
- You need speech-to-text models you can choose between by size and architecture
- You want a streaming or real-time recogniser rather than file transcription alone
- You are comparing speech recognition checkpoints from the NeMo ecosystem
