AI BriefingHugging Face Releases AnySpeechRecognizer and AnyDecisionModel Swift Packages for Apple Platforms
·2026.10.29 09:00
Key point
The packages wrap Apple's new SpeechAnalyzer API with NVIDIA and Alibaba models, achieving 1.55% WER on LibriSpeech test-clean, and enable typed decision outputs via six backends including MLX and llama.cpp.
Details
- Hugging Face released AnySpeechRecognizer and AnyDecisionModel, two Swift packages enabling local speech recognition and decision-making on Apple platforms using open-source models.
- AnySpeechRecognizer wraps Apple's SpeechAnalyzer API (macOS 26, iOS 26) to support NVIDIA Parakeet and Alibaba Qwen3-ASR engines, achieving 1.55% WER on LibriSpeech test-clean with Parakeet v3 file transcription.
- AnyDecisionModel provides typed probability outputs for yes/no, multiple-choice, and scale questions via six backends, including MLX, Core AI, llama.cpp, Jev, and Ollama.
- Speaker diarization is supported via NVIDIA Streaming Sortformer, tracking up to 4 speakers, a feature not available in Apple's native API.
- Performance on M4 Max Mac Studio shows Parakeet processing 1 hour of audio in 20-40 seconds (file) with ~100 MB memory, while Qwen3-ASR requires several GB and is slower, suitable for file transcription.
- AnyDecisionModel's next-token backends limit choices to 26 options and score levels to 10, whereas decision-head backends like Cloudflare's Clef have no such limits.
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.