Two new Swift packages let Apple-platform developers run specialized speech and decision models behind familiar interfaces. AnySpeechRecognizer mirrors Apple’s speech APIs while allowing applications to choose engines such as Nvidia Parakeet or Alibaba Qwen3-ASR running locally through Core ML.
The speech package supports transcription from files or streams and can add speaker diarization, which labels who spoke when. Because it follows the structure of Apple’s Speech framework, developers can keep much of their existing code and change the model configuration rather than rebuilding the entire pipeline.
AnyDecisionModel handles yes-or-no questions, multiple-choice selections and rating scales. Instead of asking a language model to generate prose and then parsing it, the package returns typed results and probabilities derived from model outputs. It can use on-device runtimes including MLX, Core AI and llama.cpp, or connect to decision models from Jev and Ollama.
The release argues for using task-specific models when a full language model adds cost and variability without adding value. On-device execution can also reduce network dependence and keep audio local, though developers still need to test accuracy, download size, battery use and hardware compatibility across the Apple devices they support.