Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms

Hugging Face has announced two open-source Swift packages, written up by developer Mattt on October 29, 2026, that extend its AnyLanguageModel approach to speech recognition and decision-making on Apple devices. Both can fetch open models from the Hugging Face Hub and run them locally, and both require Swift 6.2 and macOS 26 or iOS 26 or later.

AnySpeechRecognizer mirrors Apple's SpeechAnalyzer API but lets developers swap the engine, offering NVIDIA's Parakeet, Moondream's Parakeet Ultra and Alibaba's Qwen3-ASR through Core ML. It also adds speaker diarization, absent from Apple's framework, using NVIDIA's Streaming Sortformer to track up to four speakers. On an M4 Max Mac Studio, Hugging Face measured file-based word error rates on LibriSpeech's clean test set of 1.43% to 1.66% for the Parakeet models, against 2.12% for Apple's transcriber. Parakeet handled an hour of audio in roughly 20 to 40 seconds. Every engine exceeded 23% on overlapping meeting recordings, and the authors caution that testing covered one machine and has not yet reached iPhones.

AnyDecisionModel answers yes-or-no, multiple-choice and rating questions with typed results and probabilities rather than parsing generated text. It has six backends: MLX, MLX decision heads such as Cloudflare's Clef, Core AI, llama.cpp, Jev and Ollama. A demo app, Emoji Magnet, sorts 343 emoji against a typed phrase, taking about 10 seconds locally on an M4 Max.

The release matters because it gives Apple developers specialized on-device alternatives to general language models, which the authors describe as better, faster and cheaper for these tasks.