Whisper vs Parakeet WER: 7.4% vs 6.3% on the Open ASR Leaderboard

Parakeet TDT 0.6B v3 has a lower average word error rate (WER) than Whisper large-v3 on the Hugging Face Open ASR Leaderboard: 6.32% against 7.44%, as reported by Local AI Master on 20 June 2026. It is also much faster, at about 3,333x real time against 146x by real-time factor (RTFx). Every figure below comes from secondary write-ups dated June to October 2026, and none of those sources dates the leaderboard snapshot it used.

How much more accurate is Parakeet than Whisper large-v3?

Parakeet TDT 0.6B v3 has the lower average WER, by about 1.1 percentage points. Local AI Master reports 6.32% for Parakeet and 7.44% for Whisper large-v3 (20 June 2026). OpenWhispr reports 6.34% for Parakeet and roughly 7.4% for Whisper (18 July 2026). Measured against Whisper's 7.44%, Parakeet's 6.32% is about 15% lower.

WER counts the share of words a model gets wrong, so a lower number is better. A 1.1-point gap on an average of several test sets is a real difference in the headline figure. It is not a verdict on every audio type, as the sections below explain.

How much faster is each model?

Parakeet is about 23 times faster by RTFx. Local AI Master reports about 3,333x real time for Parakeet, and OpenWhispr reports 3,332x. Local AI Zone reports 146x for Whisper large-v3 on 6 October 2026. The Hugging Face model card lists 145.51 for the same model, which rounds to 146. Dividing 3,333 by 146 gives 22.8, which rounds to 23.

RTFx measures seconds of audio processed per second of compute, so a higher number means faster transcription. The sources do not state the hardware behind each value, and the Parakeet and Whisper figures come from different write-ups. The 23x ratio is therefore a comparison of reported throughput, not a controlled speed test on identical machines. If you deploy on a specific GPU, run both models there.

How big are the models, and how many languages do they cover?

Parakeet is less than half the size of Whisper large-v3, and it covers far fewer languages. Parakeet TDT 0.6B v3 has 600 million parameters and supports 25 European languages, according to Local AI Master and OpenWhispr. Whisper large-v3 has 1.55 billion parameters, as its Hugging Face model card lists it, and supports 99 languages.

Local AI Master also describes Parakeet as using a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. That architecture detail rests on that one source. The size difference matters for deployment: a 0.6 billion parameter model needs less memory than a 1.55 billion parameter model, though neither source reports memory use.

What does the Open ASR Leaderboard measure?

The leaderboard runs each model through the same English test sets in one harness, which is what makes the averages comparable across models. Local AI Zone describes the setup as eight identical English test sets run in one standardized harness (6 October 2026). The average in this post comes from those English sets.

Two metrics appear in every source. WER is the share of words a model gets wrong, so lower is better. RTFx is throughput, so higher is better. Neither metric captures punctuation, latency to first token, or behaviour on silence, and this post reports none of those for either model. A team choosing between the two should not treat the average as a proxy for those properties.

Do the sources agree on Parakeet's score?

Nearly. The two write-ups differ by 0.02 points on Parakeet's average: 6.32% from Local AI Master (20 June 2026) and 6.34% from OpenWhispr (18 July 2026). A gap that small does not change the ranking. It does mean a post that quotes one figure should name its source and date instead of presenting the number as a fixed leaderboard value.

The dates matter as well. Two write-ups published a month apart may reflect different leaderboard versions. That is an inference, since neither source states which version it used, and the gap between them is small enough that it may be unrelated.

What do the averages hide?

An average can hide a dataset where the ranking reverses. This post reports only the averages across the leaderboard's English test sets. A team choosing a model for one domain should read the per-dataset table on the leaderboard before deciding, because the average can favour one model while a single test set favours the other.

Language coverage is a separate gap. The comparison covers Parakeet's 25 European languages, and nothing in these sources says how Parakeet performs outside them. Whisper covers 99 languages, and an English average says nothing about those languages.

What's still unclear?

• Every figure is secondary. No leaderboard table is cited directly. The Parakeet and Whisper averages come from Local AI Master and OpenWhispr, and the Whisper RTFx comes from Local AI Zone, and the Hugging Face model card agrees to within rounding. The per-dataset numbers and the exact snapshot date are not in this post.
• Parakeet's average is 6.32% or 6.34%. The two sources differ, and the post cites both.
• The throughput ratio is indicative. About 23x combines figures from three write-ups, with no shared hardware.
• The weakest point of the argument is the average itself. A 1.1-point gap on English test sets says little about accent, background noise or domain. It also says nothing about the languages Parakeet does not cover.