Swift devs swap Apple speech for local Parakeet and Qwen3-ASR models

Hugging Face has released AnySpeechRecognizer and AnyDecisionModel, two new Swift packages that extend local AI capabilities on Apple platforms.confirmed According to a blog post published by Hugging Face, these tools allow developers to integrate open-source speech and decision models using familiar Apple APIs.reported

The evidence

The AnySpeechRecognizer package mirrors Apple’s SpeechAnalyzer, SpeechTranscriber and AssetInventory interfaces. It defaults to Apple’s engine but allows swapping in NVIDIA’s Parakeet or Alibaba’s Qwen3-ASR.reported Both models run locally on Core ML. The package adds speaker diarization to track who speaks when.

Parakeet v3 is NVIDIA’s TDT 0.6B v3 model. It recognises 25 European languages and auto-detects the spoken language. Moondream’s Parakeet Ultra offers higher accuracy in Hugging Face’s measurements. Parakeet Unified is an English-only variant. All three run on a custom Core ML runtime supporting live input and word timing.

Qwen3-ASR comes in 0.6B and 1.7B sizes from Alibaba. It recognises 30 languages and 22 Chinese dialects, including Mandarin, Cantonese, Japanese and Korean. It detects language automatically or accepts manual settings. It outputs one final result per 40-second segment, lacking volatile results or word-level confidence scores.

The decision

Developers can change the speech engine by altering how they create the transcriber object. The steps for installing assets and reading results remain identical to Apple’s framework. This abstraction layer means existing code using Apple’s Speech framework requires minimal modification to adopt open models.

AnyDecisionModel handles yes-or-no, multiple-choice and rating-scale questions. It derives answers from model probabilities rather than generating text. It supports models running on device via MLX, Core AI or llama.cpp, or remote decision models from Jev or Ollama. Both packages can download open models directly from the Hugging Face Hub.

What the evidence means

Parakeet models are faster and support live input with volatile results and word timing. They are better suited for real-time applications. Qwen3-ASR runs on the GPU and is slower, making it more appropriate for transcribing files than live audio streams. It lacks word timing and confidence metrics.

Adding a new engine requires implementing SpeechRecognitionModel for configuration and SpeechRecognitionExecutor for session management. Hugging Face provides a placeholder provider in the README to assist with integration. The published Core ML models pin source checkpoints and file checksums to fixed revisions for reproducibility.confirmed

What builders can do with it

Developers can build support apps that route calls based on content using AnyDecisionModel. They can implement local transcription for 25 European languages with Parakeet or 30 languages plus Chinese dialects with Qwen3-ASR. Builders can add speaker diarization to track multiple speakers in a call.

Teams can swap Apple’s default engine for open models without rewriting asset management or result-handling code. Developers can run decision models locally via MLX or llama.cpp for privacy-sensitive yes/no queries. Builders can integrate live transcription with word-level confidence using Parakeet’s Core ML runtime.