AnySpeechRecognizer and AnyDecisionModel brought open speech and decision models to Apple apps

Hugging Face released two Swift packages that let Apple apps pick their own speech and decision models, and run them on the device. The details come from the company's own launch post.

The idea starts with a support call. First you need to know what the customer said. Then you need to decide where to send them. A language model can do both, but the post says purpose-built models are better, faster and cheaper.

AnySpeechRecognizer

Apple's SpeechAnalyzer, introduced in macOS 26 and iOS 26, turns speech into text, but only Apple's engine runs behind it. AnySpeechRecognizer keeps the same API and lets developers swap the engine. Apple's own stays the default.

The engines on offer:

• Parakeet v3 from NVIDIA: 25 European languages, detected automatically.
• Parakeet Ultra from Moondream: a post-trained version of v3 that was more accurate in most of Hugging Face's measurements.
• Parakeet Unified: English only.
• Qwen3-ASR from Alibaba: it covers 30 languages plus 22 Chinese dialects, among them Mandarin, Cantonese, Japanese and Korean.

Qwen3-ASR gives one final result per segment of up to 40 seconds, with no word timing. It is slower, so it suits files better than live audio.

The package also adds speaker diarization, which means tracking who spoke when. It uses NVIDIA's Streaming Sortformer, follows up to four speakers, and copes with overlapping talk.

Speed and size

On an M4 Max Mac Studio, Parakeet handled an hour of audio from a file in roughly 20 to 40 seconds. Live, it took about 1.5 to 2.5 minutes. Each Parakeet download is about 600 MB. Qwen3-ASR needs 0.9 to 2.4 GB.

AnyDecisionModel

This package answers yes-or-no, multiple-choice and rating questions. It reads the model's probabilities instead of generating text, so the app sees how likely each option was. It works with local models through MLX, Core AI or llama.cpp, or with decision models from Jev or Ollama.

The demo app, Emoji Magnet, asks one yes-or-no question for each of 343 emoji when you pause typing. With the hosted Jev backend, the emoji settle in about half a second. Locally, Qwen3 4B on MLX takes about 10 seconds for a full pass.

The package also runs Cloudflare's Clef-Flash, a 9B model with a trained decision head. The 4-bit version is a 6.2 GB download, and 16 GB of memory is recommended.

Limits

The speech numbers come from one machine and a few datasets, and testing so far covers macOS 27 only, not iPhone. On AMI meeting recordings, every engine had a word error rate above 23%. The Clef backend takes only text and JSON, not images or video. Confidence values show how concentrated the probabilities are, not whether an answer is right.