A drop-in replacement for Apple’s transcriber
Hugging Face published two Swift packages on 29 October for Apple platforms. AnySpeechRecognizer mirrors Apple’s own speech APIs — SpeechAnalyzer, SpeechTranscriber and others — but lets a developer choose the engine behind them: Apple’s default, NVIDIA’s Parakeet, or Alibaba’s Qwen3-ASR, all running locally. It adds a SpeakerDiarizer for attributing speech to speakers, and ships a hear command-line tool.
The second package, AnyDecisionModel, answers yes-or-no, multiple-choice and rating-scale questions with typed results and probabilities. It reads the model’s own probabilities rather than parsing generated text, which removes a class of failure that comes from a model answering in prose when a caller wanted a value.
The word error rates
The author, writing as mattt, published a word error rate table measured on LibriSpeech using files of three to eight minutes, on an M4 Max running macOS 27.0.1. On test-clean, Parakeet Unified in file mode scored 1.43% and in live mode 1.45%. Parakeet v3 scored 1.55% in file mode, Parakeet Ultra 1.66%. Apple’s own transcriber scored 2.12% in file mode.
On the harder test-other split the order holds: Parakeet Unified at 2.19% in file mode, Parakeet Ultra at 2.55%, Parakeet v3 at 2.78%, and Apple’s transcriber at 3.97%. The post gives no figures for Qwen3-ASR or for Apple’s live mode, so neither can be placed in that ranking.
These are one developer’s measurements on one machine, and the post says as much. They are not an independent evaluation, and the packages’ author has an interest in the swap being worthwhile.
Where all of them fall over
The more useful number in the post is the one nobody wins. On AMI meeting recordings, where speakers talk over each other, every engine tested came in above a 23% word error rate. That is the gap between reading an audiobook and transcribing a meeting, and it has not closed.
Diarization tracks up to four speakers, including when they overlap, using NVIDIA’s Sortformer models, and runs alongside an Apple or Parakeet transcriber.
Licences and footprint
The models carry different terms. Parakeet TDT v3 and Parakeet Ultra are 0.6B parameters under CC BY 4.0; Parakeet Unified EN is 0.6B under the NVIDIA Open Model License; Qwen3-ASR comes in 0.6B and 1.7B under Apache 2.0. Sortformer 4spk v2 is CC BY 4.0 and v2.1 is under the NVIDIA licence. The Clef-Flash decision head is 9B, with no licence stated in the post.
Downloads run to about 600MB per Parakeet model, around 590MB for the live encoder, 0.9 to 2.4GB for Qwen3-ASR and about 250MB for Sortformer. The 4-bit Clef build is 6.2GB, and its model card recommends at least 16GB of memory.
The practical catches
Both packages need Swift 6.2 or later and macOS 26 or iOS 26 or later. Parakeet, Sortformer and MLX require Apple silicon. The Core AI backend needs Xcode 27 and an OS 27 runtime, and the MLX build tasks need Xcode’s Swift 6.4 or later with the Metal toolchain. AnyDecisionModel’s CoreAI trait cannot be combined with AnySpeechRecognizer at all, because of a swift-huggingface version conflict.
The post also notes that testing covered only macOS 27, not an iPhone — which is a meaningful gap for packages whose stated target includes iOS.
What to watch
The interesting thing here is not the benchmark margin, which is small in absolute terms. It is that a system API Apple ships can now be backed by an open model from a different vendor, on the user’s own device, with the application code unchanged. Whether that stays true through Apple’s next OS release is the thing to follow.
