Hugging Face releases open speech and decision models for Apple platforms

Hugging Face has released two new open-source Swift packages designed to bring specialized machine learning capabilities directly to Apple devices. Announced on October 29, 2026, AnySpeechRecognizer and AnyDecisionModel allow developers to run advanced speech-to-text engines and structured decision models locally on macOS 26 and iOS 26. These tools extend the ecosystem established by previous Swift integrations, offering alternatives to cloud-based APIs and general-purpose language models for specific tasks.

What happened

The release introduces AnySpeechRecognizer, a package that implements Apple’s new SpeechAnalyzer API while allowing developers to swap out the underlying recognition engine. Instead of relying solely on Apple’s proprietary framework, users can choose from NVIDIA’s Parakeet models or Alibaba’s Qwen3-ASR. The package also adds speaker diarization, a feature absent from Apple’s native tools, enabling apps to identify who is speaking in multi-person conversations. All models run locally using Core ML, ensuring data privacy and offline capability.

Simultaneously, Hugging Face launched AnyDecisionModel, which addresses the need for structured outputs from AI systems. Rather than prompting a large language model to generate text and then parsing it, this package uses model probabilities to answer yes-or-no questions, multiple-choice queries, and rating scales. It supports various backends, including MLX for local execution on Apple silicon, as well as remote options like Jev and Ollama. This approach provides typed results with confidence scores, making it easier to integrate AI decisions into strict software logic without fragile text parsing.

How it works

AnySpeechRecognizer mirrors the structure of Apple’s Speech framework, using modules like SpeechTranscriber and AssetInventory. Developers install models via the Hugging Face Hub, with assets downloaded only when requested. The package supports NVIDIA’s Parakeet v3, a polyglot model recognizing 25 European languages, and Qwen3-ASR, which covers 30 languages including Mandarin and Cantonese. For speaker identification, it integrates NVIDIA’s Streaming Sortformer, which tracks up to four speakers simultaneously. The Core ML conversion allows these models to run efficiently on device, with Parakeet processing an hour of audio in roughly 20 to 40 seconds from a file.

AnyDecisionModel operates by treating decisions as probability distributions over defined labels. When a developer defines a choice, such as selecting a support department, the model calculates the likelihood of each option in a single forward pass. For local execution, it uses MLX or llama.cpp backends, reading next-token probabilities or utilizing trained decision heads like Cloudflare’s Clef. To improve accuracy, the package includes features like rotation debiasing, which averages probabilities across different option orders, and calibration tools to adjust for model overconfidence. This method yields structured data types immediately, avoiding the latency and error rates associated with text generation and parsing.

Key details

• AnySpeechRecognizer supports NVIDIA Parakeet v3, Parakeet Ultra, and Alibaba Qwen3-ASR, with automatic language detection for most models.
• Speaker diarization is handled by NVIDIA’s Streaming Sortformer, capable of tracking up to four overlapping speakers in real-time.
• AnyDecisionModel provides typed outputs for binary choices, multiple-choice selections, and rating scales, complete with probability and confidence metrics.
• Local execution relies on MLX, Core AI, or llama.cpp backends, while remote options include Jev and Ollama servers.
• Benchmarks on an M4 Max Mac Studio showed Parakeet Unified achieving a 1.43% word error rate on clean speech tests.
• The packages require Swift 6.2 or later and are compatible with macOS 26 and iOS 26 or newer.

Why it matters

For engineers building on Apple platforms, these packages offer a significant shift toward modular, local AI. By decoupling the API interface from the underlying model, developers can upgrade recognition engines or decision logic without rewriting application code. This flexibility is crucial for maintaining performance and accuracy as open-source models evolve. Furthermore, running these models locally reduces dependency on external APIs, lowering costs and ensuring compliance with strict data privacy requirements that often prohibit sending audio or sensitive user data to the cloud.

The introduction of specialized decision models also addresses a common pain point in AI integration: reliability. General-purpose language models often struggle with consistent formatting, requiring complex post-processing to extract actionable data. AnyDecisionModel’s use of raw probabilities provides a deterministic structure that fits naturally into Swift’s type system. This allows developers to build more robust features, such as automated ticket routing or sentiment analysis, with measurable confidence levels rather than hoping for correct text output.

What you can do

• Install AnySpeechRecognizer and AnyDecisionModel via Swift Package Manager in projects targeting macOS 26 or iOS 26.
• Test Parakeet v3 and Qwen3-ASR against your specific audio datasets using the included hear command-line evaluation tool.
• Implement speaker diarization in meeting apps by integrating the SpeakerDiarizer module alongside your transcriber.
• Replace text-parsing logic with AnyDecisionModel for features like urgency detection or category classification to get typed results.
• Experiment with the Emoji Magnet example app to understand how batched decision queries work with MLX backends.
• Calibrate your decision models using separate datasets to adjust temperature and bias settings for optimal accuracy in production.