Apple’s speech API now takes Parakeet, at 1.43% errors on LibriSpeech

Hugging Face has released two Swift packages for Apple platforms, AnySpeechRecognizer and AnyDecisionModel. The first mirrors Apple’s SpeechAnalyzer API but lets developers pick the engine, such as Nvidia’s Parakeet, according to the official post.

One import swap changes the speech engine

AnySpeechRecognizer copies the shape of Apple’s speech classes in macOS 26 and iOS 26, including SpeechAnalyzer, SpeechTranscriber, DictationTranscriber, SpeechDetector and AssetInventory. Apple’s own engine stays the default. Code that sticks to the shared API needs only import Speech replaced with import AnySpeechRecognizer. As the post puts it, “AnySpeechRecognizer keeps the API and lets you choose the engine.”

The package ships three Parakeet engines on a Core ML runtime: Nvidia’s multilingual Parakeet TDT 0.6B v3 (25 European languages, automatic language detection), Parakeet Ultra, a post-trained variant from Moondream, and an English-only Parakeet Unified. They support file and live input, volatile results, word timing and word confidence. Alibaba’s Qwen3-ASR (0.6B and 1.7B) covers 30 languages and 22 Chinese dialects, but runs on the GPU, returns one final result per segment of up to 40 seconds and suits files better than live audio.

A SpeakerDiarizer adds speaker labels using Nvidia’s Streaming Sortformer, tracking up to four speakers, including overlap. Developers can also write their own engine by implementing SpeechRecognitionModel and SpeechRecognitionExecutor.

On LibriSpeech files of 3 to 8 minutes, run on an M4 Max Mac Studio, the author reports word error rates (test-clean) of 1.43% for Parakeet Unified, 1.55% for Parakeet v3 and 2.12% for Apple’s transcriber. Parakeet is said to process an hour of audio in roughly 20 to 40 seconds from a file. Live mode is slower and less accurate: Parakeet v3 reached 4.73% on test-clean live.

Answers drawn from probabilities, not generated text

AnyDecisionModel handles classification-style questions: yes/no, multiple choice and rating scales. A DecisionSession holds a piece of text or JSON, and methods such as probability(of:), choice(_:from:) and score(_:levels:) return typed values with probabilities. In the post’s words, it “derives each answer from the model’s probabilities instead of generating text.” Each question is independent, so earlier answers never feed later ones.

Six backends are supported: MLX, MLX models with trained decision heads, Core AI, llama.cpp, TypeSafe’s Jev and Ollama. Cloudflare’s Clef-Flash, a 9B model with a decision head, returns probabilities for all options in a single pass; its 4-bit MLX build is a 6.2 GB download, and the model card recommends at least 16 GB of memory.

The limits are spelled out. Next-token backends allow at most 26 options per choice and 10 levels per score. Confidence values come from entropy and show how concentrated the probabilities are, not whether the answer is right, so developers should measure error rates on their own data. An optional rotationDebiasing setting averages over shuffled option orders to reduce position bias, at one prompt evaluation per rotation.

Who can use it now, and what to check first

Both packages need Swift 6.2 or later and macOS 26 or iOS 26. Parakeet, Sortformer and MLX require Apple silicon. The Core AI backend needs Xcode 27 and an OS 27 runtime, and its trait cannot currently be combined with AnySpeechRecognizer because of a conflicting swift-huggingface version requirement. The sample Package.swift pins AnySpeechRecognizer from 0.1.0 and AnyDecisionModel from 0.6.0, which signals early versions. The post does not state a license for either package.

Models download only when an app asks for them, through the standard Hugging Face cache. Sizes are about 600 MB per Parakeet, 0.9 to 2.4 GB for Qwen3-ASR and about 250 MB for Sortformer. Licenses vary: the Core ML conversions of Parakeet v3 and Parakeet Ultra are CC BY 4.0, Parakeet Unified falls under the NVIDIA Open Model License, and Qwen3-ASR is Apache 2.0. Commercial teams should read those terms per model.

For developers, the practical change is that on-device transcription no longer has to mean Apple’s engine alone. Hugging Face’s earlier AnyLanguageModel 1.0 applied the same pattern to Apple’s Foundation Models framework, with local backends such as llama.cpp, MLX and Core ML and hosted ones including OpenAI, Anthropic and Gemini. The new packages extend that approach to speech and classification. Nvidia’s speech work is also moving fast: our report on Nvidia’s prompt-driven open speech model noted a 62% cut in medical transcription errors after tuning, a sign that open models can be adapted to specialist vocabularies.

Several things remain unverified. All benchmarks are the author’s own, from one machine and a few datasets, on macOS 27 only, with no iPhone results yet. The post says more engines are planned, including Core AI on macOS 27 and iOS 27, but gives no dates.