AnySpeechRecognizer and AnyDecisionModel Bring Open AI Models to Apple Swift
Two open Swift packages bring local speech recognition, speaker diarization and probability-based decisions to Apple devices with interchangeable models.
Summary
Announced October 29, 2026, AnySpeechRecognizer 0.1.0 mirrors Apple’s macOS 26 and iOS 26 SpeechAnalyzer API while adding interchangeable engines and Core ML speaker diarization. Apple remains the default; NVIDIA’s 0.6B Parakeet variants cover up to 25 European languages, while Alibaba’s 0.6B and 1.7B Qwen3-ASR cover 30 languages and 22 Chinese dialects. Qwen processes segments up to 40 seconds but lacks live partial results, word timing and confidence. NVIDIA Streaming Sortformer tracks four overlapping speakers. Hugging Face downloads are about 600 MB per Parakeet, another 590 MB for its faster live encoder, 0.9 to 2.4 GB for Qwen and 250 MB for Sortformer.
On an M4 Max Mac Studio running macOS 27.0.1, Parakeet transcribed one hour in 20 to 40 seconds from files or 1.5 to 2.5 minutes live, using about 100 MB after compilation. LibriSpeech word error rates ranged from 1.43% to 4.73%, versus Apple’s 2.12% and 3.97%; all tested engines exceeded 23% on overlapping AMI meeting audio. Testing has not reached iPhone.
AnyDecisionModel 0.6.0 returns typed yes or no, multiple-choice and rating results with probabilities instead of generated text. Its six backends are MLX, MLX decision heads, Core AI, llama.cpp, TypeSafe Jev and Ollama. Cloudflare’s 9B Clef-Flash uses a trained decision head; its 4-bit build is 6.2 GB and recommends 16 GB memory. Both packages require Swift 6.2 and macOS 26 or iOS 26; several local runtimes require Apple silicon. Core AI needs Xcode 27 and OS 27, while a swift-huggingface dependency conflict currently prevents combining AnyDecisionModel’s CoreAI trait with AnySpeechRecognizer.
Positives
- Parakeet file transcription processed one hour of audio in 20 to 40 seconds on an M4 Max Mac Studio.
- Parakeet Unified achieved 1.43% and 2.19% LibriSpeech word error rates, beating Apple’s 2.12% and 3.97%.
- Qwen3-ASR supports 30 languages and 22 Chinese dialects through downloadable 0.6B and 1.7B Core ML models.
- Streaming Sortformer identifies up to four speakers, including overlapping speech that Apple’s Speech framework cannot diarize.
- AnyDecisionModel exposes typed choices, probability distributions and confidence values across six local or hosted backends.
- Open Hugging Face models download only when requested and are pinned to fixed revisions with file checksums.
Risks & concerns
- Every tested recognition engine exceeded 23% word error on AMI meetings containing overlapping speech.
- Qwen3-ASR is slower than Parakeet, needs several gigabytes of memory and lacks partial results, word timing and confidence.
- Testing covers macOS 27 only, leaving iPhone performance and behavior unverified.
- Model downloads range from about 250 MB to 2.4 GB, while Clef-Flash requires 6.2 GB and recommends 16 GB memory.
- A swift-huggingface version conflict prevents AnyDecisionModel’s CoreAI trait from being combined with AnySpeechRecognizer.
- Next-token decision backends permit only 26 choice options or 10 score levels and may require costly rotation debiasing.