Hugging Face releases Swift speech and decision model packages

Hugging Face released two open Swift packages on October 29 for specialized AI tasks on Apple platforms. AnySpeechRecognizer handles configurable speech transcription; AnyDecisionModel returns typed model decisions across supported local and hosted backends.

The packages extend Swift model tooling for Apple platforms with separate interfaces for speech and decision tasks. Developers can switch between supported engines while keeping more application-facing code consistent, though capabilities and requirements vary by backend.

AnySpeechRecognizer follows the shape of Apple’s newer SpeechAnalyzer API family and defaults to Apple’s Speech framework. It also supports Parakeet model conversions derived from NVIDIA and Moondream work through a Core ML runtime, plus Core ML conversions of Alibaba’s Qwen3-ASR 0.6B and 1.7B models. According to the project, its SpeakerDiarizer can use NVIDIA Streaming Sortformer conversions to track up to four speakers on the transcription timeline.

That mirroring offers source-level compatibility, not interchangeable types. The repository says the package’s modules do not conform to Apple’s Speech.SpeechModule protocol, its analyzer cannot accept Apple module instances, and some Apple features remain unavailable. Apple and Parakeet can produce live, changing transcripts. Qwen3-ASR instead returns final results by segment, without interim results, word timing or confidence values. DictationTranscriber and SpeechDetector work only with the Apple backend.

The speech package manifest sets Swift 6.2 and deployment targets of macOS 26 and iOS 26. The project says Parakeet and Sortformer require Apple silicon. Hugging Face says Parakeet processed an hour of audio in about 20 to 40 seconds from a file and 1.5 to 2.5 minutes in live mode in tests on an M4 Max Mac Studio running macOS 27.0.1. These are not independent benchmarks: the announcement calls the figures specific to the tested machine and data and advises developers to test their intended hardware and audio.

AnyDecisionModel serves a different purpose. It returns typed yes-or-no probabilities, choices from a set of options, or ordinal scores. Choice and Score results include a probability distribution and a confidence value, and choices can map to Swift enum cases. For the MLX next-token backend, the project says the package prompts a model with one-token labels for the allowed answers and extracts their probabilities in a single forward pass. A separate MLX backend can read a trained decision head.

The package lists six backends: MLX next-token models, MLX models with trained decision heads, Core AI, llama.cpp with local GGUF files, TypeSafe Jev or compatible System One servers, and Ollama System One models running locally or through Ollama cloud. With no package traits enabled, the core library includes Jev and Ollama support and has no third-party dependencies, according to the repository. The MLX, CoreAI and Llama traits add the optional local engines.

The package manifest likewise sets Swift 6.2 and macOS 26 and iOS 26 deployment targets. The project also documents Linux support for the core library and the Jev and Ollama backends. MLX requires Apple silicon and the Metal toolchain. Core AI requires Xcode 27 to build and macOS 27 or iOS 27 to run, while llama.cpp requires a supported chat template in the GGUF model. The repository specifies Ollama 0.35.0 or later, and 0.40.5 or later for cloud models.

The decision outputs have not been independently evaluated for calibration. The project warns that raw language-model probabilities can be overconfident and that choice order can influence answers. Its confidence measure describes the concentration of the returned distribution, not the likelihood that an answer is correct. The project recommends measuring error rates on application data before setting thresholds. Both repositories display Apache-2.0 licenses; the speech project’s downloadable model conversions retain separate, model-specific licenses.