The Pulse
Hugging Face Adds Open Speech and Decision Models to Apple’s Swift APIs
Hugging Face announced AnySpeechRecognizer and AnyDecisionModel, Swift packages that let developers select speech engines and get typed, probability-based decisions on Apple devices. Its Mac tests show Parakeet beating Apple’s transcriber o
AI.info Team · Written with AI · fact-checked ·
Apple’s SpeechAnalyzer gives developers a speech-to-text API, but its recognition engine is Apple’s alone. Hugging Face’s answer, announced October 29, is AnySpeechRecognizer, a Swift package that preserves the API pattern while adding selectable engines, including local NVIDIA and Alibaba models. The launch also includes AnyDecisionModel, a separate package that returns typed choices and probabilities rather than a generated answer.
AnySpeechRecognizer keeps Apple’s API shape
On macOS 26 and iOS 26, Apple’s SpeechAnalyzer accepts audio files or streams and delivers results asynchronously. AnySpeechRecognizer mirrors that interface, so developers using the shared API can switch by changing the imported package; selecting another engine requires changing how they create the transcriber. Apple’s framework remains the default option.
The alternatives cover different needs. NVIDIA’s Parakeet v3 recognizes 25 European languages, while Alibaba’s Qwen3-ASR supports 30 languages and 22 Chinese dialects, including Japanese and Korean. Qwen returns a final result for audio segments up to 40 seconds, but does not provide partial results, word timing or confidence scores; Hugging Face says it is slower and better suited to files than live audio.
For conversations, the package adds NVIDIA’s Streaming Sortformer to assign words to speakers. Hugging Face says the diarizer can track up to four people, including when they speak over one another. Model files download when requested: the post lists roughly 600 MB for each Parakeet model, 0.9 to 2.4 GB for Qwen3-ASR and about 250 MB for Sortformer.
Parakeet leads in the reported Mac test
Hugging Face tested the speech package on an M4 Max Mac Studio running macOS 27.0.1, using LibriSpeech files between three and eight minutes long. Parakeet Unified recorded word error rates of 1.43% on the test-clean set and 2.19% on test-other; Apple’s transcriber scored 2.12% and 3.97%, respectively. Those are results from the package’s creator, not an independent comparison.
The limits matter alongside the favorable scores. Hugging Face says the measurements cover one computer and a few datasets, and that every tested engine exceeded a 23% word error rate on AMI meeting recordings with overlapping speech. The team has tested only on macOS 27, not iPhone; it says it wants to add more engines, including Core AI support on macOS 27 and iOS 27.
AnyDecisionModel returns probabilities, not prose
The second package targets a different gap: apps often ask a language model to classify a ticket, choose a department or rate a customer’s frustration, then have to parse the model’s text. AnyDecisionModel instead exposes yes-or-no probabilities, typed multiple-choice answers and ratings on ordered scales. “There are models built for each that are better, faster, and less expensive,” Mattt writes in the announcement.
AnyDecisionModel offers six backends: MLX, MLX decision heads, Core AI, llama.cpp, Jev and Ollama. Most can run locally, while Jev and Ollama can connect to hosted services. For a local decision-head example, the package supports Cloudflare’s Clef-Flash; its four-bit model is a 6.2 GB download, and the model card recommends a Mac with at least 16 GB of memory.
The probability output is not a guarantee of correctness. Hugging Face warns that language models can be overconfident and that a reported confidence value measures how concentrated the probabilities are, not whether an answer is right. Some next-token backends also cap choices at 26 options and ratings at 10 levels, while correcting for option-order bias requires extra evaluations.
One dependency conflict still separates the packages
Both packages require Swift 6.2 or later and macOS 26 or iOS 26 or later. The Parakeet and Sortformer runtimes and the MLX backend require Apple silicon; Core AI requires Xcode 27 and an OS 27 runtime. A separate version mismatch currently prevents developers from combining AnyDecisionModel’s CoreAI trait with AnySpeechRecognizer: their dependencies require incompatible major versions of swift-huggingface.
The release therefore offers Apple-platform developers more choice at two different stages—transcribing audio and making structured decisions—but does not make every model interchangeable in every project. For now, the clearest performance evidence is limited to Hugging Face’s Mac speech tests, while decision-model users must validate probability calibration against their own data.