Hugging Face brings Parakeet and Qwen3-ASR to iPhone and Mac

AnySpeechRecognizer keeps SpeechAnalyzer's APIs and lets you choose between Apple's engine, Parakeet and Qwen3-ASR 1

Parakeet models cover 25 European languages; Qwen3-ASR covers 30 languages and 22 Chinese dialects 1

AnyDecisionModel returns answers with probabilities instead of generated text 1

AI-generated content reviewed by the editorial team. How we work

Synthesis of 6 sources · 3 background articles ✓ Official source Background from our archive

Shared origin: All information comes from Hugging Face's announcement.

Original document: Hugging Face Blog

Sep 16, 2026 Apple's AI arrives in the macOS 27 terminal
Oct 27, 2026 Ollama brings local speech generation to Macs with Phonon
Oct 30, 2026 ServiceNow opens a 27-billion-parameter voice model
Oct 30, 2026 Hugging Face brings Parakeet and Qwen3-ASR to iPhone and Mac

Image: dev.to — https://dev.to/madnessbrains_i/hugging-face-anyspeechrecognizer-swap-apple-parakeet-or-qwen3-asr-without-rewriting-your-stt-path-479j

An app that listens to a customer first has to understand what they say, then decide where to route them. For the first step, there are models built specifically for speech recognition, faster and cheaper than a general-purpose language model. Hugging Face has released two Swift packages, AnySpeechRecognizer and AnyDecisionModel, that bring these open models to Apple devices.

The first package follows the APIs of SpeechAnalyzer, the new transcription system introduced with macOS 26 and iOS 26. By default, it uses Apple's Speech framework, but it lets developers choose another engine. To switch, you change the import line and, to use Parakeet, change only how the transcriber is created. The rest of the code, from installing models to reading results, stays the same.

The supported engines include NVIDIA's Parakeet models and Alibaba's Qwen3-ASR, in several sizes. Parakeet v3 recognizes 25 European languages and works out on its own which one is being heard. Parakeet Ultra, from Moondream, was more accurate in most of Hugging Face's measurements. Qwen3-ASR covers 30 languages and 22 Chinese dialects, including Cantonese, Japanese and Korean, which Parakeet does not handle. It is slower and runs on the GPU, so it suits transcribing files more than live listening.

The models have been converted to Core ML and published on Hugging Face under CC BY 4.0 for Parakeet and Apache 2.0 for Qwen3-ASR. On Mac Apple Silicon, Ollama has chosen MLX as the inference engine starting with version 0.40.0-rc3 in October. They download only when the app asks: about 600 MB per Parakeet, from 0.9 to 2.4 GB for Qwen3-ASR. The package also adds speaker diarization, meaning it can tell who is speaking at each moment, which Apple's Speech framework does not offer.

The second package, AnyDecisionModel, answers yes-or-no, multiple-choice and rating-scale questions. It does not generate text: it returns typed results with probabilities, computed by decision models running locally with MLX, Core AI or llama.cpp, or through Jev or Ollama.

According to Hugging Face, combining the two packages lets developers use specialized models for each task instead of handing everything to one language model. The code is available on the author's GitHub profile.

Related Articles