NVIDIA NeMo / parakeet-tdt-0.6b-v3
Parakeet TDT 0.6B v3
Ultra-fast batch transcription
Languages
25 European
Parameters
600M
License
CC BY 4.0
OpenVox fit
Fast batch STT
Model overview
Built into OpenVox for private local speech transcription and subtitles.
Parakeet TDT 0.6B v3 is NVIDIA’s high-performance multilingual ASR model built on FastConformer and Token-and-Duration Transducer (TDT) decoding for ultra-fast transcription and precise subtitle timestamps.
Achieves an impressive 6.34% mean WER on the Open ASR Leaderboard, outperforming larger models on English meeting and clean speech benchmarks.
Uses Token-and-Duration Transducer (TDT) decoding to predict both tokens and durations simultaneously, achieving major inference speedups.
Officially supports 25 European languages with native word-level and segment-level timestamp extraction for pinpoint subtitle alignment.
In OpenVox, Parakeet TDT is the recommended engine for high-volume batch audio processing and rapid subtitle creation.
All speech transcription runs 100% locally on your computer with offline model weights. No microphone audio, meeting recordings, or transcripts are ever sent to external cloud servers.
Open-source model: NVIDIA NeMo / parakeet-tdt-0.6b-v3
View on Hugging Face