Direct answer — Is Parakeet or Whisper better for file transcription? On accuracy, Parakeet is narrowly ahead on both English tracks of the 2026 Open ASR Leaderboard: 6.32% average word error rate against 7.44% for Whisper large-v3 on short-form clips, and 10.7% against 11.2% on long-form recordings. On the five European languages both models are benchmarked in, they tie at 4.81%. The decisive difference is coverage: Whisper handles 99 languages, Parakeet 25.
Every parakeet vs whisper transcription comparison published this year quotes the same headline: NVIDIA’s model is more accurate and vastly faster than OpenAI’s. That claim holds up — and it settles much less than it looks like it settles.
The Open ASR Leaderboard paper behind it runs several separate tracks, and the parakeet vs whisper accuracy gap behaves differently in each. Read all of them and the decision for a real archive stops turning on the leading decimal at all.
This guide takes them in the order that matters when you are choosing the best local transcription model 2026 for interviews, lectures and video files: what each model is, what the short-form gap measures, what happens to it on hour-long recordings, where language coverage draws a hard line, and what survives once the model leaves a datacentre GPU for your Mac.
What are Parakeet and Whisper, and how do they differ?
Two families of open speech recognition models, built on different architectures for different priorities. Neither is a product — both are weights you run inside something else.
Parakeet TDT 0.6B v3 is NVIDIA’s multilingual model, described on its card as a 600-million-parameter system with a FastConformer-TDT architecture, released under the CC-BY-4.0 licence. TDT stands for token-and-duration transducer: instead of scoring every audio frame, the decoder predicts how long to skip ahead, which is where its throughput comes from.
Whisper large-v3 is OpenAI’s encoder–decoder transformer, trained on what its card calls “1 million hours of weakly labeled audio and 4 million hours of pseudo-labeled audio collected using Whisper large-v2”, and reported to deliver “10% to 20% reduction of errors compared to Whisper large-v2”.
Whisper large-v3-turbo is the variant most Mac apps actually ship. Its card is blunt about the trade: “it’s the exact same model, except that the number of decoding layers have reduced from 32 to 4”, so “the model is way faster, at the expense of a minor quality degradation”.
| Parakeet TDT 0.6B v3 | Whisper large-v3 | Whisper large-v3-turbo | |
|---|---|---|---|
| Publisher | NVIDIA | OpenAI | OpenAI |
| Architecture | FastConformer-TDT | Encoder–decoder transformer | Same, decoder pruned to 4 layers |
| Parameters | 600 million | 1,550 M | 809 M |
| Languages | 25 (all European) | 99 | 99 |
| Punctuation and casing | Automatic | Automatic | Automatic |
| Timestamps | Word and segment level | Segment and word level | Segment and word level |
| Licence stated on the card | CC-BY-4.0 | Apache 2.0 | MIT |
Does Parakeet beat Whisper on accuracy?
On the English short-form track, yes, by 1.12 points. The Open ASR Leaderboard paper — which compares 86 open-source and proprietary systems across 12 datasets — reports an average word error rate of 6.32% for Parakeet TDT 0.6B v3 against 7.44% for Whisper large-v3.
The periphery matters more than the gap. That average is computed over the leaderboard’s English short-form datasets — including AMI, GigaSpeech, LibriSpeech clean and other, SPGISpeech, VoxPopuli and a five-hour subset of Earnings22 — which are largely pre-segmented clips rather than continuous hour-long recordings.
The speed figures come from the same run. Inverse real-time factor, defined in the paper as total audio duration divided by transcription time, was measured on an NVIDIA A100-SXM4-80GB GPU at batch size 64.
| English short-form track | Average WER | RTFx (A100, batch 64) |
|---|---|---|
| NVIDIA Parakeet TDT 0.6B v2 | 6.05% | 3,390 |
| NVIDIA Parakeet TDT 0.6B v3 | 6.32% | 3,330 |
| NVIDIA Parakeet CTC 1.1B | 7.40% | 2,730 |
| OpenAI Whisper large-v3 | 7.44% | 146 |
Read the right-hand column before repeating it: it describes a datacentre GPU processing 64 files at once, not a laptop working through one recording. Nearly every published parakeet vs whisper accuracy claim stops at this table.
Does that gap hold on hour-long recordings?
The order holds. The numbers do not. The leaderboard runs a separate English long-form track over CORAAL, Earnings21, the full Earnings22 and TED-LIUM v3 — continuous recordings rather than clips — and this is the table that matches what you drop into a transcription app.
Parakeet TDT 0.6B v3 still leads both Whisper builds there, at a 10.7% average word error rate against 11.0% for large-v3-turbo and 11.2% for large-v3. But its 1.12-point short-form lead shrinks to half a point, and every open model in the table roughly doubles the error rate it posted on clips.
| Long-form track (open models) | Average WER | RTFx (A100, batch 64) |
|---|---|---|
| Cohere Labs Transcribe | 9.73% | 418 |
| NVIDIA Parakeet TDT 0.6B v3 | 10.7% | 1,000 |
| OpenAI Whisper large-v3-turbo | 11.0% | 148 |
| OpenAI Whisper large-v3 | 11.2% | 68.6 |
| Distil-Whisper large-v3.5 | 11.7% | 156 |
| NVIDIA Parakeet CTC 1.1B | 12.9% | 2,790 |
Three things are worth taking from it. Neither family tops the track — Cohere Labs Transcribe leads the open models, and four proprietary API systems sit above all of them, from 7.32% to 9.54%. Parakeet stays dramatically faster throughout, so its throughput advantage is real even where its accuracy lead is marginal. And continuous audio is simply harder for everyone: a 6% benchmark model is an 11% model on a two-hour recording, whichever badge it carries.
The same Earnings22 corpus makes the point. The short-form track uses a five-hour subset of it; the long-form track uses all 119 hours, and the error rates move accordingly.
The architectural reason is visible on NVIDIA’s own card. Parakeet supports “audio up to 24 minutes long with full attention (on A100 80GB) or up to 3 hours with local attention” — a documented ceiling that a two-hour interview can meet, and a mode switch that a transcription app has to handle for you.
Which model covers more languages for file transcription?
Whisper, by a factor of roughly four, and this is the one place where the two models genuinely separate. Parakeet TDT 0.6B v3 supports 25 languages, all European; OpenAI’s cards for large-v3 and large-v3-turbo both state 99.
Inside the overlap they are indistinguishable. The leaderboard’s multilingual track benchmarks five languages — German, French, Italian, Spanish and Portuguese, over the CoVoST-2 and FLEURS datasets — and both models land on exactly the same average.
| Average WER, multilingual track | Whisper large-v3 | Parakeet TDT 0.6B v3 |
|---|---|---|
| German | 4.26% | 4.20% |
| French | 6.36% | 5.42% |
| Italian | 4.69% | 4.81% |
| Spanish | 3.65% | 3.73% |
| Portuguese | 4.96% | 6.16% |
| Average of the five | 4.81% | 4.81% |
A tie, split across the languages: Parakeet takes German and French, Whisper takes Italian, Spanish and Portuguese. Outside those five it is not a comparison at all. Japanese, Arabic, Hindi, Mandarin, Korean, Turkish and Hebrew have no Parakeet language token, and no amount of throughput compensates for a language the model was never trained to emit.
That boundary, not half a point of word error rate, is why Weesper Transcribe is built on Whisper rather than on the faster model: 57 transcription languages decide whether a file in a mixed archive gets transcribed at all.
What changes when the model leaves the A100 and runs on your Mac?
Almost everything about speed, and very little about accuracy. This is the step where leaderboard numbers stop being predictive.
Three shifts happen at once:
- Batching disappears. An RTFx of 3,330 assumes 64 files processed in parallel on an 80 GB datacentre GPU. You transcribe one file, on a laptop.
- The weights get quantised. Mac apps ship compressed models to fit memory and battery budgets, which changes both the speed and the error rate away from the published figure.
- The runtime decides the ceiling. whisper.cpp, the plain C/C++ implementation of Whisper, ships Core ML and Metal support and quantised model files — which is precisely why so much Mac tooling is built on Whisper rather than on the faster model.
Weesper Transcribe takes that route: Whisper on-device through whisper.cpp, a quantised turbo model alongside a lighter base model, with Metal acceleration on Apple Silicon and support for Intel Macs with AVX2, on macOS 13 or later. Nothing is uploaded, so an interview under embargo stays on the machine that recorded it.
For the model-level detail behind that choice, our comparison of turbo against Distil-Whisper covers what pruning and distillation each cost, and the wider survey of open-source speech models sets both against the newer entrants.
How should you choose between them for files?
Answer four questions in order and the best local transcription model 2026 for your own material usually picks itself.
- What language is the recording in? Outside Parakeet’s 25 European languages, the comparison ends here. Whisper is the only one of the two that has a token for the language.
- How long are the files? Both families lose roughly four points of accuracy on continuous recordings compared with segmented clips, so budget correction time on a two-hour file whichever model produced it.
- What hardware runs it? Parakeet’s throughput was measured on an A100 at batch 64. If your workload is one file at a time on a Mac, the published ratio is not the ratio you will see.
- What do you need after the transcript? Timestamps, subtitles, batch queues, corrections and search are properties of the application, not of the weights.
That last point is the one benchmark tables cannot express. On long-form audio the two families sit half a point apart, while the gap between processing a two-hour recording in one pass and slicing it by hand is the difference between a finished job and a lost afternoon.
In practice the workflow decides: batch processing across a folder, nine export formats including time-coded SRT and VTT, inline editing with clickable timestamps, a local automation API, and a history you can search by meaning rather than by keyword.
Conclusion
Parakeet TDT 0.6B v3 wins the benchmark most articles quote, and it wins the one they skip as well: ahead of Whisper on short English clips and on hour-long recordings alike, with throughput no encoder–decoder model comes near. On the five European languages both are measured in, the two tie exactly at 4.81%.
Which leaves the choice where the tables cannot reach it. Half a point of word error rate on long-form audio is not a decision; 99 supported languages against 25 is. Beyond Parakeet’s European set there is no second candidate, and on a Mac the runtime, the batch queue and the long-file handling settle the rest.
Want to test that on your own recording? See how on-device file transcription works, or get the app from the Mac App Store — free to download, macOS 13 or later, files up to 15 minutes at full quality, with a one-time Pro upgrade rather than a subscription.