Hugging Face released two open-source Swift packages on October 29, 2026, expanding its platform tools beyond language generation to speech recognition and probability-based decision-making on Apple operating systems.
The packages, AnySpeechRecognizer and AnyDecisionModel, provide Swift interfaces that connect local and hosted machine learning models to Apple devices. Both packages can download open models directly from the Hugging Face Hub, requiring Swift 6.2 or later alongside macOS 26 or iOS 26 or later.
Speech recognition and diarization
AnySpeechRecognizer implements the SpeechAnalyzer architecture that Apple introduced in macOS 26 and iOS 26, but allows developers to swap Apple's default recognition engine for open alternatives running on Core ML. The package mirrors Apple's SpeechTranscriber, DictationTranscriber, SpeechDetector, and AssetInventory types.
Hugging Face converted several speech models to Core ML for the release, including NVIDIA's Parakeet TDT 0.6B v3, Moondream's post-trained Parakeet Ultra, and NVIDIA's English-only Parakeet Unified. The package also supports Alibaba's Qwen3-ASR in 0.6-billion and 1.7-billion parameter sizes, covering 30 languages and 22 Chinese dialects. Qwen3-ASR runs on the graphics processing unit without intermediate word timings, making it better suited to pre-recorded files than live streams.
Because Apple's framework lacks a speaker diarization module, AnySpeechRecognizer introduces SpeakerDiarizer. The module runs NVIDIA's Streaming Sortformer v2 or v2.1 to track up to four speakers simultaneously, assigning timestamps to words even when speakers talk over one another.
Hugging Face evaluated the models on an M4 Max Mac Studio running macOS 27.0.1. Parakeet v3 transcribed an hour of audio in 20 to 40 seconds from a file, recording a 1.55% word error rate on LibriSpeech benchmarks compared to 2.12% for Apple's built-in transcriber. Live transcription with Parakeet took 1.5 to 2.5 minutes per audio hour, operating with approximately 100 megabytes of memory once compiled by Core ML.
Decision models
AnyDecisionModel handles categorical tasks, including yes-or-no questions, multiple-choice selections, and rating scales. Rather than prompting a large language model to generate text, the package extracts answers and confidence values directly from the model's output probabilities in a single forward pass.
The framework operates across six backends. Local execution relies on Apple silicon using MLX, Core AI on macOS 27 and iOS 27, or llama.cpp via llama.swift. Hosted options include TypeSafe's Jev service and local or cloud instances running Ollama 0.35.0 or later. The package also supports models equipped with dedicated decision heads, such as the 9-billion parameter Clef-Flash running on MLX.
To counter common classification biases, the local backends support temperature and bias calibrations as well as option-rotation debiasing. A package conflict currently prevents developers from combining AnySpeechRecognizer with the CoreAI trait of AnyDecisionModel, because coreai-kit 0.7.3 depends on swift-huggingface versions below 1.0, while the speech package requires version 1.x.
