On October 29, developer Mattt introduced AnySpeechRecognizer and AnyDecisionModel on Hugging Face. The former offers a choice of speech-recognition engine; the latter returns typed binary, multiple-choice and rating-scale answers with probabilities. These are independent Swift projects, rather than an Apple announcement of a new model.
A structured answer is not necessarily a correct decision
AnyDecisionModel's README describes a session with immutable text or JSON context. Questions are independent: one answer does not become input for the next. Execution can be local or hosted: Jev is remote, while Ollama can be local or cloud-based. CoreAI requires Xcode 27 and OS 27.
Our editorial assessment is that typed results simplify integration with ordinary program logic. An application can handle a predefined choice rather than searching for a freely generated word in an answer. However, a structured format does not eliminate misclassification or turn a returned probability into measured reliability for a particular business process.
For example, a personal-notes organiser could suggest a folder from a defined set. It should allow correction, particularly when a note combines several subjects. If the result triggers a more consequential action, the automation threshold and opportunity for human intervention require a separate decision. This is an illustrative use case, not a package test.
Similar interfaces, different limitations
AnySpeechRecognizer follows the modern SpeechAnalyzer interface, offering Apple, Parakeet and Qwen3-ASR engines. Interface similarity does not mean every engine supports identical features.
The current README states that Qwen3-ASR processes files without word timestamps or a context prompt, and has not been tested on iOS devices. Physical iPhone, OS 26 and minimum Swift 6.2 testing remain outstanding for Parakeet. A dependency conflict between AnyDecisionModel's CoreAI integration and AnySpeechRecognizer is also documented. Package code uses Apache 2.0; model weights have separate licences.
For developers, the practical boundary is the particular combination of platform, version and engine. A recording button working with one backend does not justify promising the same behaviour, speed and extra attributes with another. Similarly, using Swift does not guarantee that data stays on the device: that depends on the selected execution path.
What has been checked
We read the announcement and both public projects' documentation, but did not compile them or test them on devices. The development is a software layer enabling model choice and clearer result types. README limitations are part of evaluating that choice, especially when promising cloud-free operation or compatibility across several backends.
*AI illustration: conceptual Swift applications, not a screen from an actual product.*
*Editorial verification: 29.10.2026, 19:17:12 +0200 (Europe/Sofia).*
