AnySpeechRecognizer and AnyDecisionModel

Hugging Face Swift packages: swappable speech engines behind Apple's SpeechAnalyzer, and typed probability answers from local models.

Hugging Face · Developer tools and coding agents · open-source

Published in Feed

Overview

Hugging Face introduced AnySpeechRecognizer and AnyDecisionModel on October 29, 2026, two Swift packages that follow the pattern of AnyLanguageModel. AnySpeechRecognizer adds engine choice to Apple's SpeechAnalyzer API (macOS 26 or iOS 26): Apple's engine stays the default, with NVIDIA Parakeet v3, Parakeet Ultra, Parakeet Unified and Alibaba Qwen3-ASR as alternatives, plus a Streaming Sortformer speaker diarizer for up to four speakers. AnyDecisionModel answers yes/no, multiple-choice and rating questions with typed results and probabilities, with MLX, Core AI, llama.cpp, Jev and Ollama backends.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Switching the import is often enough to swap the SpeechAnalyzer engine; Core ML conversions are pinned to fixed Hub revisions
  • Parakeet Unified scored 1.43% word error rate on LibriSpeech test-clean versus 2.12% for Apple's transcriber (one M4 Max, macOS 27 only, authors' test)
  • Qwen3-ASR 0.6B and 1.7B cover 30 languages and 22 Chinese dialects but lack word timing and confidence
  • SpeakerDiarizer based on NVIDIA Streaming Sortformer v2 or v2.1 attributes timed words to up to four speakers
  • AnyDecisionModel derives probabilities from token probabilities, with calibration, rotation debiasing and entropy-based confidence; caps at 26 choices and 10 score levels

Use cases

  • Swift apps that need on-device transcription with a more accurate model than Apple's default
  • Meeting transcription with speaker attribution on macOS
  • Local classification or filtering in an app, such as the Emoji Magnet example that filters 343 emoji from a typed phrase

Pricing

Free and open source; model licenses vary (CC BY 4.0, Apache 2.0 for Qwen3-ASR, NVIDIA Open Model License).

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
29 Oct 2026
Record updated
3 Nov 2026
Record last checked
3 Nov 2026

Explore the Feed

Other entries filed under developer tools and coding agents; inclusion does not imply equivalent features.

  • Claude Skills Marketplace

    Anthropic's public marketplace for Claude Agent Skills, letting developers publish, monetize, and install reviewed Skills directly into Claude Code and the Claude apps.

  • Composer 2

    Cursor's second-generation in-house coding model, now the default agent model in the editor, with faster multi-file edits and native terminal tool use.

  • Devin 3

    Cognition's autonomous software engineering agent, now able to run unsupervised multi-day engineering sprints with persistent repo memory and automatic PR-review handling.

Browse the Feed