How to choose an ASR model for your Mac

An automatic speech recognition model, or ASR model, is the part of a transcription app that turns audio into words. Everything else (speaker labels, summaries, subtitles) is built on its output. When the model runs on your own Mac, the recording never has to leave it. The catch is that you have to pick one, and the model names do not tell you much about which fits your work.

Speech recognition models in TranscribeX at a glance

WER is the average word error rate on the Hugging Face Open ASR Leaderboard’s English short-form sets (lower is better), from the leaderboard paper (Srivastav et al., arXiv 2510.06961) or the model’s own card where the paper has no entry. Sizes and languages from the OpenAI Whisper README and model cards, the NVIDIA Parakeet, Nemotron and Canary model cards on Hugging Face, the Qwen3-ASR model cards (the 0.6B figure is from its Hugging Face evaluation panel), and Apple’s WWDC25 session 277. The Nemotron WER is for NVIDIA’s English streaming model; the multilingual model is scored on FLEURS instead. Apple publishes no WER or parameter count. Checked October 2026.

Our test: five models on an M4 Pro Mac

Leaderboard numbers come from datacenter GPUs. To see how the models behave on a Mac, we ran the same 8-minute English recording through each model that was installed in TranscribeX on our test machine.

Our own test, run October 17, 2026 on a Mac with an Apple M4 Pro chip and 24 GB of memory, macOS 27.0.1. Audio: 73 clean read-speech clips from the LibriSpeech validation set (hf-internal-testing/librispeech_asr_dummy on Hugging Face), joined into one 8-minute file, 1,150 reference words. Each model was run twice through the TranscribeX command line and the second run is shown; times are end to end, including loading the model. Word error rate uses simple normalization (lowercase, punctuation removed). Clean read speech is easier than a real meeting, so expect higher error rates on your own calls, and treat differences under one point as noise on a sample this size. Qwen3 ASR was not timed in this round; Whisper Large V3 Turbo, Parakeet V3 and Nemotron were not part of it.

Two things stand out. On clean English speech, Apple’s free built-in model was as accurate as Parakeet V2 and about as fast as Whisper Tiny, which makes it the best free choice for supported languages. Whisper Large V3 was accurate but by far the slowest; on a Mac, Parakeet gives you the same accuracy in a fraction of the time for English. Whisper still matters for the 100 or so languages the others don’t cover.

NVIDIA Canary is in the table because it comes up in almost every ASR model comparison. TranscribeX does not run it. The original canary-1b is licensed for non-commercial use only, and Canary-Qwen 2.5B was trained on audio clips up to 40 seconds, so it is a poor fit for hour-long meetings without extra chunking.

What the numbers mean (and what they don’t)

Word error rate (WER) counts the words a model got wrong (substituted, dropped or invented) divided by the words in the reference transcript. A WER of 6% means roughly six mistakes per hundred words. The leaderboard strips punctuation and casing before scoring, so its numbers are not directly comparable with figures a vendor publishes on its own test setup.

RTFx is how many seconds of audio a model transcribes per second of compute. On the leaderboard, Parakeet V2 reaches an RTFx of about 3,380 and Whisper Large V3 about 146. Those runs happen on datacenter GPUs with large batches, so they tell you how models rank against each other, not how long a file will take on your MacBook. Use them to compare, then test on your own audio.

The differences between the top models on English benchmarks are small: 6.05% for Parakeet V2 and 6.42% for Qwen3 ASR 0.6B against 7.44% for Whisper Large V3. On a real recording, microphone quality, crosstalk and accents usually matter more than a point of WER. A model that handles your language well and finishes fast enough that you actually use it is the right best ASR model for you.

How to choose: four questions

• What language is the audio in? English only: Parakeet V2. One of 25 European languages: Parakeet V3. Chinese, Japanese, Korean, Thai, Vietnamese, Arabic or Hindi: Qwen3 ASR or Whisper. Anything else, or a mix you can’t predict: Whisper Large V3.
• Do you need the text while the audio is still playing? Then you need a streaming model. In TranscribeX that is Nemotron for live captions and live meeting transcripts, or the low-latency Parakeet EOU model for English. Apple On-Device can also stream.
• How much do you transcribe? If it is a few files a week, Whisper Large V3’s slower speed won’t matter. If you batch hours of recordings, Parakeet’s speed adds up quickly.
• Do you need exact timing? For subtitles or karaoke-style highlighting, Qwen3 ASR with its forced aligner is built for this: Qwen reports an average alignment error of 42.9 ms, against about 130 ms for WhisperX and the NeMo Forced Aligner.

Free vs Pro models

The free version of TranscribeX transcribes without limits using Whisper Tiny and Apple On-Device. Whisper Tiny is small and quick, and good enough for clear audio and notes to yourself. Apple’s model is a strong option for its 30 supported locales and costs nothing to try. Pro ($39 once) unlocks the rest of the Whisper family, fine-tuned Whisper models for languages like German, Swiss German, Korean and Hebrew, Parakeet, Qwen3 ASR and Nemotron.

If you care more about privacy than about a particular model, all of these run locally. TranscribeX also offers ElevenLabs Scribe and Mistral Voxtral as optional online models; they send audio to those providers, so leave them off if the recording must stay on your Mac. More on that in offline transcription on a Mac.

Which ASR model for which job

Recommendations are based on the published figures in the first table and each model’s language list. They are a starting point, not a guarantee for your audio.

You can switch models per file in TranscribeX, so the simplest test is to run the same recording through two of them and read the results side by side. If you are coming from MacWhisper, many of the same models are available, and this comparison covers how the apps differ.

Model guides