Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms
Web page · https://huggingface.co/blog/anydecisionmodel
First seen by Oriel on .
AI summary
The article introduces AnySpeechRecognizer and AnyDecisionModel, Swift packages that run selectable speech-recognition and decision models locally on Apple platforms through familiar APIs.
- AnySpeechRecognizer mirrors Apple's SpeechAnalyzer, SpeechTranscriber, and AssetInventory APIs while adding Parakeet, Qwen3-ASR, and speaker diarization through Sortformer.
- Parakeet runs on a custom Core ML runtime with file and live input, timing, and confidence; it transcribes an hour in 20–40 seconds from a file.
- Qwen3-ASR supports 30 languages and 22 Chinese dialects, but produces only final segment results, runs slower on GPU, and suits files better than live input.
- AnyDecisionModel returns typed choices, scores, and probabilities by reading allowed-answer probabilities rather than generated text, using independent questions over text or JSON state.
- Backends include MLX, Core AI, llama.cpp, Jev, and Ollama; calibration and rotation debiasing address overconfidence and option-position bias, but confidence is not correctness.
#swift #speech recognition #core ml #decision models
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/anydecisionmodel | https://huggingface.co/blog/anydecisionmodel Read discussion |