Whisper large-v3-turbo, explained.

What OpenAI changed

• The announcement says turbo performs similarly to large-v2 across languages, with larger degradation on some languages like Thai and Cantonese. It does better on FLEURS, which has cleaner recordings, than on Common Voice.
• Multilingual only. OpenAI's table has no English-only turbo build, and the language list in the repo's tokenizer.py runs to 100 entries.
• MIT license. OpenAI's README says Whisper's code and model weights are all released under it.

A Whisper maintainer announced turbo in the repo's discussions on 1 October 2024. It has only 4 decoder layers, the same as the tiny model, down from the 32 in the large series. The announcement credits the Distil-Whisper paper, which found that a smaller decoder greatly improves speed with minimal loss of accuracy. Turbo was not distilled, though. It was fine-tuned for two more epochs over the same multilingual transcription data used for large-v3, excluding translation data.

The model card calls it the exact same model apart from the decoding layers. So the encoder, which listens, keeps its full size, and the decoder, which writes the text out token by token, is one eighth of its old depth.

Run turbo with the openai-whisper package

Since package version 20240930 the whisper command line defaults to turbo, so the --model flag below is only there to be explicit.

pip install -U openai-whisper

# transcribe with turbo
whisper audio.wav --model turbo

# translate non-English speech into English: not turbo
whisper japanese.wav --model medium --language Japanese --task translate

Run turbo with whisper.cpp

whisper.cpp is a plain C/C++ port. Its README lists CPU-only inference and calls Apple Silicon a first-class citizen, so it suits a laptop with no NVIDIA GPU. It offers turbo in ggml format, with q5_0 and q8_0 quantized versions if you want the file smaller. These are the project's own quick start steps with the model name swapped.

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh large-v3-turbo

cmake -B build
cmake --build build -j --config Release

./build/bin/whisper-cli -m models/ggml-large-v3-turbo.bin -f samples/jfk.wav

Run turbo with faster-whisper

faster-whisper runs Whisper through CTranslate2. Both large-v3-turbo and the short name turbo are in its model list, and it fetches the converted weights from the Hugging Face Hub on first use. Note that segments is a generator: nothing runs until you iterate.

pip install faster-whisper

# then, in Python:
from faster_whisper import WhisperModel

model = WhisperModel("large-v3-turbo", device="cuda", compute_type="float16")
# or on CPU:
# model = WhisperModel("large-v3-turbo", device="cpu", compute_type="int8")

segments, info = model.transcribe("audio.mp3", beam_size=5)
for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

How turbo scored in our own tests

• On clean read English all four are within one point, and the 95% confidence intervals of turbo, Apple and Parakeet overlap. Treat it as a tie. LibriSpeech is also in Parakeet's training data, while Whisper is scored zero-shot.
• On real conference call audio, Parakeet had the lowest error and turbo the highest, but the whole spread is about one point. Parakeet and the cloud engine have overlapping confidence intervals. Our benchmark notes record the Open ASR Leaderboard listing the reference turbo build at 11.07 on this set on 2026-07-15, level with Parakeet, so our Core ML figure likely undersells turbo.
• Outside English, turbo was ahead of Parakeet on all four FLEURS languages we ran on both: German 5.33 against 6.21, Spanish 5.68 against 6.05, French 6.82 against 7.83, Russian 7.48 against 9.18. Each set is 200 clips, so small gaps are not firm.
• Parakeet v3 covers 25 European languages and does not do Japanese, Korean or Chinese at all. Turbo scored 5.11 on Japanese and 6.01 on Chinese, both character error rate, and 13.66 on Korean. The cloud engine had the lowest error on every non-English set, ahead of turbo by about half a point to four points.
• Parakeet v3 ran at about 0.02 times real time on the test Mac, so about 50 times faster than the audio.

These are our own runs, not vendor figures. Word error rate in percent, lower is better. The Fraze row is Parakeet TDT 0.6B v3, the model the app ships.

Read the turbo row with care. Its two English scores came from WhisperKit's Core ML build of turbo on a Mac, and the FLEURS scores from Groq's hosted large-v3-turbo. Neither is the reference PyTorch build, and a different runtime or quantization moves the number. Method and code are in the sources.

When turbo is the right pick

• You transcribe files after the fact. Turbo covers far more languages than Parakeet and is a sensible default for batch work. Fraze is live only and does not transcribe files.
• The speech is in a language Parakeet does not cover and you want to stay on-device.
• You do not need speech translation. If you do, OpenAI's guidance is medium or large.

Fraze made a different call because it is live, not file based. A caption has to appear within about half a second to a second of a pause, on a laptop that is also running the meeting, with a translation step after it. A transducer model behind a voice activity detector fits that budget, and it held up on call audio. Fraze is $29.99 once or $2.99 a month after the first 30 free minutes, with no per-minute meter.

Questions