Hugging Face releases AnySpeechRecognizer and AnyDecisionModel for Apple platforms

AnySpeechRecognizer integrates speech-to-text and speaker diarization, supporting 25 European languages (Parakeet) or 30 languages plus 22 Chinese dialects (Qwen3-ASR). It mirrors Apple’s SpeechAnalyzer API but allows swapping engines—e.g., replacing Apple’s default with Parakeet v3 or Qwen3-ASR. The package includes a Sortformer diarizer for tracking up to four speakers in overlapping speech. Benchmarks on an M4 Max Mac Studio show Parakeet transcribing an hour of audio in 20–40 seconds (file) or 1.5–2.5 minutes (live), while Qwen3-ASR is slower but handles more languages. AnyDecisionModel provides typed yes/no, multiple-choice, or rating-scale decisions via local models (MLX, Core AI, or Ollama) or cloud services (Jev). Unlike traditional language-model prompts, it returns probabilities and confidence scores for each option, enabling structured decision-making without text parsing. The package supports models like Clef-Flash (9B parameters) for joint schema decisions or smaller models like Qwen3-4B for emoji classification (e.g., filtering emojis by user input). Both packages require Swift 6.2+ and macOS 26/iOS 26+, with Apple silicon needed for Parakeet/Sortformer. The code is open-source on GitHub, with models hosted on Hugging Face Hub and downloads triggered on-demand by the app.