AI BriefingAnySpeechRecognizer: Swift package supporting open models compatible with Apple Speech API
·2026.11.02 16:00
Key point
AnySpeechRecognizer, a Swift package that reimplements the Apple Speech API to enable the use of open models, and AnyDecisionModel for decision tasks have been released.
Details
- Mattt released the Swift packages AnySpeechRecognizer (0.1.0) and AnyDecisionModel (0.6.0) for Apple platforms via the Hugging Face blog
- AnySpeechRecognizer reimplements the Apple Speech API with an identical interface, enabling the use of open model engines such as Parakeet and Qwen3-ASR
- Parakeet v3 achieves a lower error rate than the Apple engine (2.12%) with a LibriSpeech WER of 1.55% (clean) on an M4 Max Mac
- Qwen3-ASR is the only open engine for Korean recognition, recording a CER of 3.93% on the FLEURS test based on the 1.7B model
- AnyDecisionModel reads token probabilities instead of generating sentences to respond to yes/no, multiple-choice, and scoring questions with typed values
- Supports six backends including MLX, MLX decision head, Core AI, llama.cpp, Jev, and Ollama, and provides an option to resolve choice order bias
- Model weights are available for download as Core ML conversions from the Hugging Face Hub, with licenses varying by model
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.