AnySpeechRecognizer and AnyDecisionModel packages enable local speech transcription and structured decision-making on Apple platforms

The AnySpeechRecognizer and AnyDecisionModel packages have been introduced for Apple platforms. The AnySpeechRecognizer package preserves the interface of the new speech transcription API from Apple, allows developers to switch the recognition model, and adds speaker diarization. According to the authors, code that uses the shared portion of the API only requires changing the import; selecting a different model also requires adjusting how the transcription component is created.

Parakeet and Qwen3-ASR models are available for local transcription through Core ML. The Parakeet TDT 0.6B v3 model recognizes 25 European languages and supports streaming results, word timings, and confidence scores. The Qwen3-ASR model, available in variants with 0.6 and 1.7 billion parameters, recognizes 30 languages and 22 Chinese dialects. It returns a final result for segments of up to 40 seconds, without interim results or word timings. It uses the GPU and, according to the authors, is slower than Parakeet models, making it better suited to transcribing files.

Models are downloaded only when requested by the application. Each Parakeet model requires a download of approximately 600 MB, with an additional encoder for live input requiring another 590 MB; Qwen3-ASR models require 0.9 to 2.4 GB, and the Sortformer model for speaker diarization requires approximately 250 MB. The Parakeet TDT 0.6B v3 and Parakeet Ultra models are licensed under CC BY 4.0, the Parakeet Unified EN 0.6B model under NVIDIA Open Model License, and both Qwen3-ASR variants under Apache 2.0. For the Streaming Sortformer 4spk model, version v2 is licensed under CC BY 4.0 and version v2.1 under NVIDIA Open Model License.

The AnyDecisionModel package provides typed responses and probabilities for yes/no questions, multiple-choice questions, or ratings on a scale. It derives responses from model probabilities instead of generating text. It supports local execution through MLX, Core AI, or llama.cpp, as well as decision models from the Jev or Ollama services. Both packages allow open models to be downloaded from Hugging Face Hub and run on the device.