NVIDIA Parakeet vs Whisper Large-v3

Is NVIDIA Parakeet faster and better than Whisper Large-v3 for long-form transcription?

Data / Setup

Input was a 90 minutes podcast audio.
Two speakers. Light background noise.
Both models ran on local GPU.

Model comparison

Feature	Whisper Large-v3	Parakeet TDT 0.6B v3
Languages	100+	25 EU langs (auto-detect)
Task	STT + translation	STT only
Max audio length	Unlimited (chunked)	Unlimited (chunked)
Multi-speaker / diarization	No (needs pyannote)	No (needs NeMo MSDD)
Timestamps	Word-level	Word-level
Parameters	1.5B	0.6B
License	MIT	CC-BY-4.0
Backend	OpenAI / HuggingFace	NVIDIA NeMo

Result

Model	WER	Runtime
Whisper Large-v3	5.1%	2:10
Parakeet 1.1B	4.4%	0:42

WER = Word Error Rate. Input: 90 min podcast, 2 speakers, light background noise, GPU.

Verdict

Parakeet wins on both axes for English.
Whisper still wins for multilingual content.

Further Links

• https://huggingface.co/spaces/hf-audio/open_asr_leaderboard