Parakeet v3 vs Whisper for meeting transcription.
Whisper Large was the default for local speech to text for years. NVIDIA's Parakeet TDT 0.6B v3 (open weights, Apache 2.0) now runs four times faster on the same hardware with slightly better accuracy on English meeting audio. Here is the landscape we studied before deciding how Notabium transcribes.
What changed
Parakeet uses a Transducer architecture rather than the encoder decoder approach that Whisper uses. Transducers are streaming friendly. They emit tokens as they hear audio, instead of waiting for a chunk to complete. For meeting transcription that maps cleanly onto how you actually consume the output: a live transcript that fills in as people speak.
The model is 600 million parameters. Whisper Large is 1.5 billion. Despite being smaller, the streaming architecture plus better training data lets Parakeet match Whisper Large's accuracy on English meeting audio and beat it on speed.
Real numbers, our laptop
On an M3 Max with the same one hour test meeting (two speakers, conference call audio, light overlap):
| Model | RTF (real time factor) | Accuracy (WER) | Notes |
|---|---|---|---|
| Whisper Large v3 | 0.18x (slow) | 5.4% | Best multilingual; slow on CPU; needs GPU for live |
| Whisper Medium | 0.42x | 6.9% | Good middle ground; English only |
| Parakeet TDT 0.6B v3 | 0.05x (fast) | 4.8% | English only; streaming friendly |
Lower RTF is better: 0.05x transcribes in 1/20 of real time. Whisper Large at 0.18x is fast enough for live transcription with a GPU, but Parakeet's headroom matters once you run transcription and a summary model on the same machine.
Where each one breaks
Whisper breaks when:
- The microphone signal is poor (over-compressed Zoom audio with packet loss). The model hallucinates plausible-sounding lines that nobody said.
- Multiple speakers overlap heavily. Whisper has no internal diarisation.
- You need streaming output. The encoder-decoder architecture waits for a chunk to finish.
Parakeet breaks when:
- Audio is not English. It is English only. Whisper covers 99 languages.
- The speaker has a heavy accent the training data did not see. Whisper Large covers more accent diversity.
- You need word-level timestamps with sub-second accuracy. Parakeet's timestamps are coarser.
What Notabium uses, and why
Running these models yourself means managing GPUs, model downloads, and language fallbacks. That setup tax rarely pays off, and a self-hosted model's accuracy ceiling sits below a managed service that is continuously tuned on meeting audio.
So Notabium transcribes in the cloud with ElevenLabs Scribe: a managed speech-to-text service built for multi-speaker meeting audio, with strong accuracy across accents, speaker labels, and timestamps. You get the quality bar this comparison is chasing without owning the infrastructure. Recordings and transcripts are encrypted in transit and at rest and scoped to your account, and AI summaries run with Claude.
The Parakeet-vs-Whisper debate is a great window into where open speech models are heading. For shipping a product people rely on every day, a managed cloud model wins on accuracy and on the simple fact that it just works.
Related: Zoom transcription · Bot vs no-bot architecture