Hugging Face ships Swift packages for speech and decision models
Hugging Face released AnySpeechRecognizer and AnyDecisionModel, two Swift packages with a shared API and swappable models underneath. Speech recognition supports Parakeet v3 (25 languages) and Qwen3-ASR (30 languages), while decisions return probabilities instead of generated text.
- Parakeet v3 hit 1.55% WER on LibriSpeech versus 2.12% for Apple's transcriber
- Qwen3-ASR covers 30 languages and 22 Chinese dialects but runs slower
- SpeakerDiarizer tracks up to four speakers, including overlapping speech
- Packages require Swift 6.2 and macOS 26 or iOS 26, versions 0.1.0 and 0.6.0
Read next
AI