Skip to content
Tools · Oct 29, 2026

Hugging Face releases AnySpeechRecognizer and AnyDecisionModel for on-device speech and decision tasks on Apple platforms

New Swift packages integrate open speech-to-text and decision models into Apple’s SpeechAnalyzer API, with Core ML backends and speaker diarization.

Trust79
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • Hugging Face launched AnySpeechRecognizer and AnyDecisionModel, two Swift packages for on-device speech recognition and decision-making on Apple platforms.
  • AnySpeechRecognizer integrates with Apple’s SpeechAnalyzer API and supports multiple open speech models, including NVIDIA’s Parakeet and Alibaba’s Qwen3-ASR, running via Core ML.
  • AnyDecisionModel answers structured yes/no, multiple-choice, and rating-scale questions using model probabilities, with support for MLX, Core AI, or llama.cpp backends.
  • Models are downloaded from the Hugging Face Hub and run locally; speaker diarization is included for multi-speaker conversations.
  • The packages are designed to work with macOS 26 and iOS 26 APIs and support live and file-based audio inputs.

Hugging Face announced two new Swift packages—AnySpeechRecognizer and AnyDecisionModel—designed to integrate open speech and decision models into Apple platforms using familiar APIs and on-device compute. AnySpeechRecognizer implements Apple’s SpeechAnalyzer API with support for multiple open speech-to-text models, including NVIDIA’s Parakeet (v3, Ultra, and Unified variants) and Alibaba’s Qwen3-ASR (0.6B and 1.7B sizes), all running via Core ML. The package mirrors Apple’s Speech framework interfaces, allowing developers to swap engines by changing a single import and model configuration. It also adds speaker diarization using NVIDIA’s Streaming Sortformer (v2 or v2.1), enabling attribution of transcribed words to up to four speakers in multi-party audio.

AnyDecisionModel answers structured questions—yes/no, multiple-choice, and rating-scale—by deriving answers from model probabilities rather than generating free-form text. It supports backends including MLX, Core AI, and llama.cpp, and can run models from the Hugging Face Hub or local providers like Jev or Ollama. Both packages download open models on demand from the Hugging Face Hub and run them locally, with model binaries cached in the standard Hugging Face cache directory.

The AnySpeechRecognizer package supports both file-based and live audio inputs, with Core ML conversions optimized for each use case. Parakeet models run on a custom Core ML runtime that supports volatile results, word timing, and confidence scores, while Qwen3-ASR provides finalized segment results without streaming features. Model sizes range from roughly 600 MB for Parakeet variants to 0.9–2.4 GB for Qwen3-ASR variants, with additional downloads for live-optimized encoders and diarization models.

Core ML conversions of the underlying models are published on the Hugging Face Hub under permissive licenses (CC BY 4.0 or Apache 2.0), with each repository recording the original checkpoint and checksums for reproducibility. The conversion workflows are open-source and include checks against the original models. Developers can target macOS 26 and iOS 26, which introduced the SpeechAnalyzer API, and integrate modules such as SpeechTranscriber, DictationTranscriber, SpeechDetector, and AssetInventory to manage model assets and processing pipelines.

Hugging Face notes that these packages extend its AnyLanguageModel approach to specialized tasks, emphasizing performance and privacy by keeping inference on-device. The announcement includes command-line tools and README examples to demonstrate integration and performance characteristics, with speaker diarization and multi-language support highlighted as differentiators over Apple’s built-in Speech framework.

Sources
  1. 01Hugging Face — Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms
Also on Tools

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.