AnySpeechRecognizer and AnyDecisionModel: open models for speech and decisions on Apple platforms
On October 29, 2026, two Swift packages were released: AnySpeechRecognizer mirrors the SpeechAnalyzer API but allows swapping the speech recognition engine, AnyDecisionModel covers selection tasks. Both download open models from Hugging Face Hub and run them on device.
On October 29, 2026, Mattt published an announcement of two Swift packages: AnySpeechRecognizer and AnyDecisionModel. Both are built on the same scheme as AnyLanguageModel — a familiar API plus a choice of model beneath it. Both can download open models from Hugging Face Hub (a repository of open weights and datasets) and run them on device. By this point AnyLanguageModel and swift-huggingface had reached 1.x releases, and the new announcement continues the same line.
The difference from what Apple offers is visible in the speech example. In macOS 26 and iOS 26, SpeechAnalyzer appeared — a speech recognition API. It works like this: modules such as SpeechTranscriber and SpeechDetector are added to the analyzer, a file or a stream of audio buffers is supplied as input, and results are read from async sequences as they arrive. AssetInventory is responsible for installing models, meaning the app itself determines the download moment. The API is neat, but there is only one engine beneath it — Apple's. AnySpeechRecognizer keeps the API and lets you choose the engine: the note lists mirrors of SpeechAnalyzer, SpeechTranscriber, DictationTranscriber and SpeechDetector, after which the source text breaks off.
The practical scenario from the note is a support call. First you need to understand what the person said (speech recognition), then decide where to route them (decision). Both tasks can be given to a language model, but the author claims that specialized models exist for each — better, faster and cheaper. Running on device removes network dependence, but the author does not name the cost of this in memory and speed. AnyDecisionModel covers selection tasks in the same way: the package pulls an open model and computes locally.
What exactly the supported models are, what licenses and hardware requirements they have, is not stated in the note. How the AnyDecisionModel API is structured and when to expect stable releases of both packages — there is no data.