1. A transducer versus an encoder-decoder
Parakeet TDT 0.6B v3 pairs a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder, 600 million parameters in all, published on Hugging Face on August 14, 2025 under CC BY 4.0. A transducer walks through the encoded audio and emits text aligned to it; the TDT variant also predicts how many audio frames each token covers, so it can skip ahead. The TDT paper reports up to 2.82 times faster inference than conventional transducers for speech recognition, and the FastConformer paper reports a 2.8 times speed-up over the original Conformer encoder.
Whisper is a Transformer encoder-decoder, also called a sequence-to-sequence model: the encoder reads a 30-second window of audio and the decoder generates the transcript one token at a time, conditioned on what it has already written. OpenAI trained the original models on 680,000 hours of multilingual audio, and large-v3 on 1 million hours of weakly labeled audio plus 4 million hours pseudo-labeled by large-v2. Large-v3-turbo is the same network with its decoder cut from 32 layers to 4, which its model card describes as much faster at the cost of a minor quality drop.
The training data shape the output. Parakeet v3 combines about 10,000 hours of human-transcribed speech with 660,000 hours of pseudo-labeled audio from NVIDIA’s Granary corpus, keeping punctuation and capitalization in every transcript. Whisper was trained to predict the raw text of transcripts without significant standardization. Both therefore produce punctuated, capitalized sentences rather than lowercase word strings, which is what you want when the text lands in an email or a document.
- Parakeet TDT 0.6B v3: 600M parameters, FastConformer-TDT, 25 European languages.
- Whisper large-v3: 1,550M parameters, encoder-decoder, 99 languages on its model card.
- Whisper large-v3-turbo: 809M parameters, 4 decoder layers instead of 32.
2. English results on the Open ASR Leaderboard, October 2026
The Open ASR Leaderboard runs every model through one evaluation script and reports two numbers from the same run: average word error rate (WER, lower is better) and RTFx, seconds of audio processed per second of compute (higher is faster). Its default English average, as displayed on October 3, 2026 for the version last updated October 2, covers eight public test sets plus two private sets of scripted and conversational speech.
On that view Parakeet v3 is both more accurate and roughly 8 to 13 times faster than the two Whisper checkpoints. Its English-only predecessor, Parakeet TDT 0.6B v2, scores slightly better still at 5.48% with an RTFx of 6,025, so v2 remains a sound choice if you only ever dictate English.
The average hides real variation. On AMI meeting recordings Parakeet v3 scores 9.42% against 13.63% for large-v3, and on the cleaned VoxPopuli set 3.19% against 4.53%. Whisper large-v3 wins on SPGISpeech financial audio (2.71% against 3.63%) and on the private scripted set (3.61% against 4.27%), while Parakeet leads on the private conversational set (13.44% against 15.13%).
Two caveats before reading these as dictation numbers. The leaderboard normalizes text before scoring, removing punctuation and case, so punctuation mistakes are not counted. And RTFx is measured on data-center GPUs with batching (the leaderboard moved its evaluations to NVIDIA H200 GPUs on June 24, 2026), so it ranks throughput, not the pause you notice after releasing a key on a laptop.
- Parakeet TDT 0.6B v3: 5.66% average WER, RTFx 6,076.
- Parakeet TDT 0.6B v2 (English only): 5.48%, RTFx 6,025.
- Whisper large-v3: 6.5%, RTFx 470.
- Whisper large-v3-turbo: 7.03%, RTFx 795.
3. Long recordings and non-English speech change the margins
The leaderboard’s long-form tab, built from earnings calls and the CORAAL sociolinguistic interviews, narrows the accuracy gap but not the speed gap: Parakeet v3 averages 13.37% WER at an RTFx of 1,003, Whisper large-v3-turbo 13.6% at 148, and Whisper large-v3 13.92% at 69. Architecture explains part of it. NVIDIA says Parakeet v3 can take up to 24 minutes of audio in one pass with full attention, or up to 3 hours with local attention, while Whisper processes 30-second windows and shifts them along the recording.
The multilingual tab averages read-speech test sets (FLEURS, CoVoST-2 and MLS) per language, and here the large Whisper model often comes out ahead. Whisper large-v3 posts lower WER than Parakeet v3 in German (4.0% vs 4.12%), Italian (4.45% vs 4.66%), Spanish (3.35% vs 3.71%), Portuguese (4.94% vs 6.08%) and Dutch (4.42% vs 7.53%); Parakeet v3 leads in French (5.38% vs 6.24%). Against large-v3-turbo, Parakeet wins in German, French, Italian and Spanish and loses in Portuguese and Dutch.
Speed still separates them: on the multilingual tab Parakeet v3’s RTFx is 3,372, against 309 for large-v3 and 550 for turbo. So outside English the trade is speed and size against a fraction of a point of WER in most languages, and several points in Dutch. That still favors Parakeet on an ordinary computer for most European languages, but it is worth a side-by-side test in yours.
4. Hallucinations, repeat loops and skipped sentences
OpenAI’s model card is direct about Whisper’s main failure: because the models were trained on large-scale noisy data, their output may include text that was never spoken, which the card calls hallucination, and the sequence-to-sequence design makes them prone to repetitive text that beam search and temperature scheduling only partly fix. The Whisper paper lists getting stuck in repeat loops, skipping the first or last words of a segment, and complete hallucination unrelated to the audio among the errors that remained as the models grew.
Pauses make it worse. In the 2024 Careless Whisper study, roughly 1% of transcriptions from OpenAI’s Whisper service, tested in 2023, contained entire hallucinated phrases or sentences, and hallucinations occurred disproportionately for speakers with longer stretches of non-vocal audio. Dictation is full of such pauses, so Whisper tools usually trim silence before transcribing: faster-whisper ships a Silero VAD filter, and whisper.cpp has a --vad option that passes only detected speech to the model.
Parakeet is not immune. NVIDIA’s technical report says it added 36,000 hours of non-speech audio to the training data specifically to reduce hallucinations inherited from Whisper-based pseudo-labels; Parakeet v3 was trained on the speech-recognition subset of that collection. The failures users report on the v3 model page look different from Whisper’s: whole sentences missing from the output (threads opened in September 2025 and May 2026) and stray words from other European languages. Read the text before you send it, whichever model produced it.
- Whisper: invented phrases and repeat loops, more likely across long silences.
- Parakeet v3: occasional dropped sentences and wrong-language words, per user reports.
- Parakeet’s card: words absent from its training vocabulary are unlikely to be recognized.
5. Streaming, latency and the hold-to-talk loop
Most desktop dictation is not streaming. You hold a key, speak, release, and the app transcribes the clip; Handy’s README describes exactly that loop. For that pattern the question is how quickly the model finishes the clip you just spoke on your own processor, which is where Parakeet’s throughput helps.
If you want words to appear while you are still talking, the architectures diverge. NVIDIA’s report notes that transducer decoders such as RNN-T and TDT are widely adopted for streaming because of their low latency, and the v3 card shows chunked streaming through a NeMo script whose example uses 2-second chunks with 10 seconds of left context. NVIDIA’s newer Parakeet Unified EN goes further: one English model for offline and streaming use, with latency as low as 160 ms. On NVIDIA’s own evaluation it scores 5.91% offline and 6.52% at 0.56 seconds of latency, while Parakeet TDT 0.6B v2 climbs from 6.04% offline to 22.83% at 1.12 seconds, so not every Parakeet is a streaming model.
Whisper was trained on 30-second chunks and has no native streaming mode. whisper.cpp’s stream example, which its README calls a naive approach, samples the microphone every half second and runs the transcription continuously. In a hold-to-talk app, Whisper simply transcribes the clip after you release the key.
6. Hardware, memory and a sensible default
NVIDIA tests Parakeet v3 on data-center and workstation GPUs from Volta to Blackwell and lists Linux as its preferred operating system, but the card also says the model needs at least 2 GB of RAM to load and points to NeMo-Speech.cpp, a native C++ runtime that runs an 8-bit GGUF build locally. Desktop apps rely on ports like these: Handy describes its Parakeet V3 option as CPU-optimized and its Whisper models as using GPU acceleration when available.
Whisper’s needs scale with size. OpenAI’s reference table runs from about 1 GB of VRAM for tiny to about 10 GB for large, with turbo at about 6 GB and roughly 8 times large’s speed on an A100. whisper.cpp brings that down with quantization and Apple silicon acceleration: it lists small at 466 MiB on disk and about 852 MB in memory and large at 2.9 GiB and about 3.9 GB, and says running the encoder through Core ML on Apple silicon is more than 3 times faster than CPU-only.
In IraVoice the default is Parakeet Unified EN, about 700 MB, built from NVIDIA’s English parakeet-unified-en-0.6b. Parakeet V2, Parakeet V3 for 25 European languages, and Whisper Small, Medium, Turbo and Large are optional downloads that also run on the computer; you can switch models under Models. IraVoice does not publish per-model speed or accuracy figures, so dictate the same paragraph with two models and keep the one that gets your names and numbers right without a noticeable wait.
- English only: start with a Parakeet model (Unified EN, TDT v2 or v3).
- German, French, Spanish, Italian and the other listed languages: Parakeet v3, then compare Whisper large-v3.
- Any other language, or one you must force: Whisper large-v3 or turbo.