New Swift packages bring selectable on-device speech recognition and typed decision models to Apple platforms

AnySpeechRecognizer provides a SpeechAnalyzer-compatible API with Apple, NVIDIA Parakeet, and Alibaba Qwen3-ASR engines, plus speaker diarization; AnyDecisionModel returns typed yes/no, multiple-choice, and rating answers with probabilities. Both packages can run open models on device, but support and performance vary: Parakeet supports live transcription, while Qwen3-ASR is slower and lacks word timing and interim results. The reported benchmarks used one M4 Max Mac Studio and a limited set of datasets.

For Apple app teams, AnySpeechRecognizer’s shared API could make switching from Apple’s engine to Parakeet a small code change, while adding diarization enables speaker-attributed transcripts. The trade-off is substantial model downloads and uneven capabilities: Qwen3-ASR needs gigabytes of memory and is better suited to files, and meeting recordings still exceeded 23% word error rate across tested engines; results may differ beyond the single tested machine and datasets.