---
type: "post"
id: "introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55"
slug: "introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55"
title: "Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms"
description: "Hugging Face has introduced AnySpeechRecognizer and AnyDecisionModel, two Swift packages for using open speech and decision models in Apple-platform applications."
retrieval_nugget: "Hugging Face has introduced AnySpeechRecognizer and AnyDecisionModel, two Swift packages for using open speech and decision models in Apple-platform applications. The release matters because it separates two common product functions—transcription and bounded classification—from general-purpose text generation while keeping model choice and local execution available to developers."
published_at: "2026-10-30"
updated_at: "2026-10-30"
record_date: "2026-10-30"
date_kind: "published_at"
topics: ["open-models"]
entities: ["Hugging Face","AnySpeechRecognizer","AnyDecisionModel","SpeechAnalyzer","Parakeet","Qwen3-ASR","Streaming Sortformer","Apple"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "Hugging Face announces two new Swift packages for speech recognition and typed decision models, extending its Apple-platform model tooling beyond AnyLanguageModel; AnyDecisionModel also incorporates Cloudflare’s Clef."
prior_refs: ["memory:640687b9-5dec-8ef0-ae9b-133f7c40a1bc","memory:033a8f0f-5035-8ad4-afa5-d8155aacc2eb"]
source_url: "https://huggingface.co/blog/anydecisionmodel"
source_urls: ["https://huggingface.co/blog/anydecisionmodel"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55/","markdown":"https://newruntime.com/posts/introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55.md","json":"https://newruntime.com/posts/introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55.json"}
---

# Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms

## Retrieval answer

Hugging Face has introduced AnySpeechRecognizer and AnyDecisionModel, two Swift packages for using open speech and decision models in Apple-platform applications. The release matters because it separates two common product functions—transcription and bounded classification—from general-purpose text generation while keeping model choice and local execution available to developers.

## Summary

Hugging Face has introduced AnySpeechRecognizer and AnyDecisionModel, two Swift packages for using open speech and decision models in Apple-platform applications. The release matters because it separates two common product functions—transcription and bounded classification—from general-purpose text generation while keeping model choice and local execution available to developers.

## What happened

In a post published on October 29, 2026, Hugging Face announced [AnySpeechRecognizer and AnyDecisionModel](https://huggingface.co/blog/anydecisionmodel), two packages extending its Swift-oriented model tooling. The stated premise is that applications often need specialist machine-learning functions alongside language models: speech recognition to establish what a caller said, and a decision mechanism to route or classify that input. Both packages can download open models from the Hugging Face Hub and run them on-device.

AnySpeechRecognizer mirrors Apple’s newer Speech API, introduced with macOS 26 and iOS 26, including SpeechAnalyzer, SpeechTranscriber, DictationTranscriber, SpeechDetector, and AssetInventory. Apple’s framework remains the default engine, but changing the import and selecting a different transcriber lets an app use NVIDIA Parakeet or Alibaba Qwen3-ASR locally. Parakeet v3 recognizes 25 European languages automatically; Parakeet Ultra is a post-trained version of v3; and Parakeet Unified is English-only. These variants run through the package’s Core ML runtime and support file and live input, volatile results, word timing, and word confidence. Qwen3-ASR is offered in 0.6B and 1.7B versions, supports 30 languages and 22 Chinese dialects, and can detect or receive a specified language. It produces final results for segments of up to 40 seconds, lacks volatile output, word timing, and confidence, and is described as slower and better suited to file transcription than live input.

The package also adds speaker diarization, which Apple’s Speech framework does not provide. Its SpeakerDiarizer runs NVIDIA Streaming Sortformer alongside an Apple or Parakeet transcriber, tracking up to four speakers even when speech overlaps. Hugging Face reports that its converted model repositories record source checkpoints and checksums and are pinned to fixed revisions. Download sizes are about 600 MB per Parakeet model plus roughly 590 MB for the live encoder, 0.9–2.4 GB for Qwen3-ASR, and about 250 MB for Sortformer. On an M4 Max Mac Studio running macOS 27.0.1, the post reports Parakeet file-transcription word error rates of 1.43%–1.66% on LibriSpeech test-clean, versus 2.12% for Apple’s transcriber. It also cautions that all tested engines exceeded 23% word error on overlapping-speech AMI meeting recordings and that the results come from one machine and limited datasets.

AnyDecisionModel addresses classification-style questions rather than transcription. A developer supplies text or JSON as a session state, then asks yes/no, multiple-choice, or rating-scale questions. The package returns typed results and probabilities rather than generated prose to parse. Its MLX backend constructs prompts with one-token labels for allowed answers and reads each label’s probability from one forward pass. Available backends include on-device MLX, Core AI, and llama.cpp, plus Jev and Ollama. The post’s example routes a support ticket, estimates urgency, and scores customer frustration; questions are independent, though they can be submitted as a batch.

## What the archive adds

This is a new event on a related open-model thread. The archive’s [Jais-ASR note](https://newruntime.com/posts/the-companion-jais-asr-notes-describe-an-open-1cbe78/) covered an open speech-recognition model focused on Arabic and deployability; AnySpeechRecognizer instead provides an Apple-facing package and multiple selectable engines. The [Open d1 note](https://newruntime.com/posts/open-d1-edge-decision-models-for-text-vision-e8468e/) covered edge decision-model checkpoints, while this release supplies a typed Swift interface for decision tasks across several backends.

## Trend signal

The release weakens the archive assumption that realtime voice systems will routinely replace chained speech-to-text and text-to-speech stacks within one model, described in [Generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/). Hugging Face’s design keeps speech recognition and decision-making as separate, swappable capabilities. Its claim is not that one voice model should own the session, but that application code should be able to choose specialized local components.

## Connections to other events

The package sits beside two previously documented developments: deployable open speech recognition in [Jais-ASR](https://newruntime.com/posts/the-companion-jais-asr-notes-describe-an-open-1cbe78/) and edge-oriented decision checkpoints in [Open d1](https://newruntime.com/posts/open-d1-edge-decision-models-for-text-vision-e8468e/). The connection is architectural rather than a shared release: this announcement provides a Swift integration layer for tasks those model categories address.

## Why it matters

For teams building Mac or iOS software, the useful change is not simply another model download. It is an API boundary that lets an application retain a familiar Apple-style speech workflow while selecting a local recognition engine according to language coverage, latency behavior, and output requirements. A team can also treat routing and scoring as bounded decisions with explicit probabilities rather than as free-form model text that must be parsed and validated afterward.

That distinction is particularly relevant for support, meeting, and workflow products. Speaker attribution and word timing can make a transcript inspectable; typed categories can make downstream routing exhaustive in Swift. The release does not remove model evaluation work, but it gives developers an implementation path in which model selection is more isolated from application logic.

## Limits and uncertainty

The post reports measurements from an M4 Max Mac Studio, macOS 27.0.1, and a limited set of datasets; it explicitly says builders should measure on their own hardware and audio. It does not establish iPhone performance, since testing so far was only on macOS 27. The published accuracy figures do not imply comparable results for noisy calls, domain vocabulary, or overlapping meetings, where every tested engine exceeded 23% word error. Nor does the announcement compare the reliability, cost, or operational trade-offs of every listed decision backend.

## Skill to build now

Build a small evaluation harness around one real workflow before choosing an engine. For speech, test representative recordings for language mix, overlap, latency, word timing, and speaker attribution; for decisions, record the allowed choices, probability distribution, and the human-correct answer. Keep the application interface stable while swapping Parakeet, Qwen3-ASR, or a decision backend, so the choice is based on measured task behavior rather than a model label.

## Reader takeaway

Hugging Face is packaging local speech transcription, diarization, and probability-based classification as replaceable Swift capabilities. Apple-platform builders can use that separation to avoid treating every product task as a text-generation prompt—but must validate language coverage, latency, and error rates on their own data.

## Open question

Will the same API and performance characteristics hold on iPhone hardware once the packages receive mobile testing?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)