Parakeet v3 vs Whisper

What each model is

NVIDIA's model card calls parakeet-tdt-0.6b-v3 a 600 million parameter multilingual model with a FastConformer encoder and a TDT decoder. It is a transducer: a large encoder turns audio into features and a small decoder predicts each token and how many frames to skip. It detects the spoken language itself. It went up on Hugging Face on 14 August 2025 under CC BY 4.0.

OpenAI's repository calls Whisper a Transformer sequence to sequence model, trained at once on recognition, speech translation and language identification. Sizes run from tiny at 39 million parameters to large at 1,550 million. The turbo checkpoint, added in September 2024, has 809 million. Its card calls it a fine-tuned version of a pruned large-v3, with the decoding layers cut from 32 to 4. Code and weights are MIT.

Both write punctuation and capital letters: NVIDIA's card lists them, and the Whisper paper says the model learns them from raw transcripts. NVIDIA lists word level and segment level timestamps. Whisper's card gives sentence level timestamps with return_timestamps set to True, and word level with it set to word.

Accuracy in our own runs

Word error rate in percent, lower is better: the share of words a model gets wrong. LibriSpeech clean is read audiobook English and Earnings-22 is real conference calls. FLEURS is 200 clips per language, measured on 30 August 2026 on a Mac mini. Parakeet ran through parakeet-mlx. Whisper turbo ran through WhisperKit, a Core ML port, for the English sets, and as a hosted copy of large-v3-turbo for FLEURS.

Whisper turbo was ahead on clean read English, 1.93 against 2.15, and on every other language we ran both on: German 5.33 against 6.21, Spanish 5.68 against 6.05, French 6.82 against 7.83, Russian 7.48 against 9.18. NVIDIA's card lists LibriSpeech among Parakeet's training data, so the clean set flatters Parakeet if anything. With 200 clips per language the Whisper figures carry an interval of more than a point either way, so read the Spanish gap as noise and the Russian one as the clearest. The cloud engine led every non-English set.

Parakeet was ahead on the earnings calls, 11.20 against 12.35, with intervals of 10.71 to 11.74 and 11.87 to 12.86. Do not read that as a verdict on Whisper. We ran a Core ML port, not OpenAI's reference code, and the Open ASR Leaderboard, which runs the reference builds, had the two 0.3 points apart on its original Earnings-22 set at the time of checking: 10.77 for Parakeet v3, 11.07 for turbo. Call them about level on this kind of audio. Parakeet's interval and the cloud engine's overlap too, so that one is a tie.

Speed and streaming

On our Mac Parakeet v3 ran at about 0.02 times real time, roughly 50 times faster than the audio plays. That run used parakeet-mlx on the GPU, not the int8 CPU files below, and we have no matching figure for Whisper. OpenAI's table only compares its own sizes on an A100 GPU, where turbo is about 8 times the speed of large. The Open ASR Leaderboard runs both models and, at the time of checking, listed throughput of about 6,000 times real time for Parakeet v3 and about 800 for Whisper large-v3-turbo.

Neither is a streaming model as shipped. Parakeet v3's encoder reads a whole clip. NVIDIA's card points to a chunked streaming script in NeMo, and sherpa-onnx files the model under offline recognizers that you feed one speech segment at a time. Whisper transcribes a sliding 30 second window and pads shorter audio to 30 seconds, and OpenAI's model card says Whisper models cannot be used for real time transcription out of the box. Live tools for either model slice the audio first.

How to run Parakeet

• The unpacked int8 files come to about 640 MB, most of it the encoder.
• That binary reads a whole file. For a microphone, the same page uses sherpa-onnx-vad-microphone-offline-asr, which puts a Silero voice activity detector in front of the same recognizer. For a long file it uses sherpa-onnx-vad-with-offline-asr.

NVIDIA's card lists three routes: NeMo, its Python toolkit, installed with pip install -U nemo_toolkit['asr'], for fine tuning and batch jobs, plus NeMo-Speech.cpp and Transformers. Outside NVIDIA, sherpa-onnx is a small C++ runtime that loads int8 ONNX files and runs on a CPU, which suits an app. These commands come from the k2-fsa documentation and assume you have built sherpa-onnx.

wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2
tar xvf sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2
rm sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2

build/bin/sherpa-onnx-offline \
  --encoder=./sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8/encoder.int8.onnx \
  --decoder=./sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8/decoder.int8.onnx \
  --joiner=./sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8/joiner.int8.onnx \
  --tokens=./sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8/tokens.txt \
  --model-type=nemo_transducer \
  ./sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8/test_wavs/en.wav

How to run Whisper

Three routes. The openai-whisper package is the reference implementation and needs ffmpeg. faster-whisper is a reimplementation on CTranslate2, the usual pick for speed in Python. whisper.cpp is a plain C and C++ implementation with no dependencies.

# whisper.cpp
git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en
cmake -B build
cmake --build build -j --config Release
./build/bin/whisper-cli -f samples/jfk.wav

# or the reference Python package
pip install -U openai-whisper
whisper audio.flac audio.mp3 audio.wav --model turbo

When Whisper is the better pick

• Your audio is not one of Parakeet v3's 25 European languages. Whisper's model card counts 99, so Japanese, Korean, Chinese, Arabic and Hindi are covered.
• You want speech translation into English from the same model. Whisper has a translate task built in, though OpenAI's README says turbo is not trained for it and returns the original language, so use medium or large.
• You need the wider ecosystem. whisper.cpp alone documents Core ML, CUDA, Vulkan and OpenVINO builds.
• You need a smaller model. Whisper comes in six sizes, from 39 million parameters up. Parakeet v3 comes in one.

Parakeet earns its place the other way round: 25 European languages in one small model, accuracy close to Whisper turbo, and speed to spare for live captions.

Questions