Open Speech and Decision Models Bring On-Device AI to Apple Platforms

On-Device Speech and Decision Models for Apple Developers

The landscape of machine learning on Apple platforms is evolving rapidly, with new tools empowering developers to integrate sophisticated AI capabilities directly into their applications. Two newly introduced Swift packages, AnySpeechRecognizer and AnyDecisionModel, aim to simplify the deployment of open speech recognition and decision-making models on iOS and macOS devices.

AnySpeechRecognizer builds upon Apple's SpeechAnalyzer API introduced in iOS 26 and macOS 26, offering developers a choice of recognition engines beyond Apple's default. It supports models such as NVIDIA's Parakeet and Alibaba's Qwen3-ASR, all running locally via Core ML. This flexibility allows apps to maintain performance and privacy while leveraging state-of-the-art speech recognition technology.

Flexible Engines and Speaker Tracking

The package mirrors Apple's native Speech framework, making integration straightforward for existing projects. Developers can switch between engines with minimal code changes, and the addition of speaker diarization helps track conversation flow in multi-speaker scenarios. This is particularly valuable for customer service applications where identifying speakers enhances context.

AnyDecisionModel complements the speech tools by providing structured responses to yes-or-no, multiple-choice, and rating-scale questions. Instead of generating text, it derives answers from model probabilities, ensuring faster and more deterministic outcomes. It supports various backends including MLX, Core AI, and llama.cpp, giving developers multiple deployment options.

Both packages can fetch open models from the Hugging Face Hub, reinforcing the trend toward accessible and transparent AI development. For organizations investing in GPU-accelerated infrastructure, these tools highlight the growing importance of efficient, on-device inference that reduces latency and dependency on cloud connectivity.