Hugging Face releases AnySpeechRecognizer and AnyDecisionModel for Apple platforms

AnySpeechRecognizer mirrors the SpeechAnalyzer API from Apple's Speech framework but lets developers pick the recognition engine. Apple's engine is the default, and switching to NVIDIA's Parakeet or Alibaba's Qwen3-ASR only requires changing how the transcriber is created. Parakeet runs on a Core ML runtime and covers 25 European languages, while Qwen3-ASR covers 30 languages and 22 Chinese dialects and runs on the GPU.

The package also adds SpeakerDiarizer, which uses NVIDIA's Streaming Sortformer to track up to four speakers and assign timed words to each one. The converted Core ML models are hosted on the Hugging Face Hub, and downloads start only when an app requests them. On LibriSpeech test-clean files, Parakeet v3 reached a 1.55% word error rate against 2.12% for Apple's transcriber, in measurements taken on one M4 Max Mac Studio.

AnyDecisionModel gets yes-or-no, multiple-choice and rating answers from the model's probabilities instead of parsing generated text. It supports MLX, Core AI, llama.cpp, Jev and Ollama backends, and one session can answer several questions about the same input. The packages require Swift 6.2 or later and macOS 26 or iOS 26, and the Parakeet, Sortformer and MLX paths need Apple silicon.

Developers building agents on Apple platforms can swap speech engines behind the SpeechAnalyzer API they already use and add speaker labels that Apple's framework does not provide. Decision calls return typed results with probabilities rather than text that has to be parsed, which makes thresholds and fallbacks easier to write. The published accuracy figures come from one machine and a few datasets, so measure on your own audio and hardware before relying on them.