Hugging Face’s New Swift Packages: Evaluate Speech and Routing Models Before Shipping

AnySpeechRecognizer mirrors Apple’s SpeechAnalyzer family of APIs while allowing developers to select a recognition engine. Apple’s Speech framework is the default; Parakeet and Qwen3-ASR provide additional options running through Core ML. AnyDecisionModel handles yes-or-no questions, multiple-choice decisions, and rating scales with typed results and probabilities.

The October 29 post by Mattt positions both packages as extensions of the familiar Swift model-and-session approach. Our assessment is that their useful contribution is a common application interface for evaluating specialized models. That does not make the engines interchangeable in every capability or establish classification accuracy.

This is source-based analysis of the announcement and repository documentation accessed November 1, 2026. We have not compiled either package or run inference. All measurements below belong to the publisher; the evaluation plan is our editorial proposal.

The announcement describes Parakeet support for file and live input, provisional results, word timing, and word confidence. Those features matter when an app must update a transcript during a conversation or attach speaker labels to individual words.

Qwen3-ASR expands the documented language coverage to 30 languages and 22 Chinese dialects. It produces final results for segments of up to 40 seconds and lacks provisional results, word timing, and confidence in this integration.

Its live-input status needs an explicit caveat. The AnySpeechRecognizer README’s Qwen section says live input works, with results trailing audio by up to 40 seconds. The same README’s Limitations section says Qwen transcribes files only. The announcement recommends file transcription. Until that discrepancy is resolved, evaluate Qwen with recorded files and treat live support as unconfirmed.

SpeakerDiarizer adds NVIDIA’s Streaming Sortformer alongside an Apple or Parakeet transcriber. The documentation describes tracking up to four speakers and assigning timed words to speakers. Accuracy degrades with five or more speakers, according to the README.

Detecting overlapping speakers does not mean recovering everything they say. The README notes that transcription usually captures the louder voice, so a brief interjection may have no corresponding words. For a meeting app, speaker attribution and transcript completeness therefore need separate acceptance checks.

The announcement reports measurements on an M4 Max Mac Studio running macOS 27.0.1. On AMI meeting recordings, every tested engine had a word error rate above 23%. These publisher measurements caution against treating clean-speech benchmarks as meeting accuracy. The announcement also says testing had covered macOS 27, not an iPhone.

AnyDecisionModel returns a typed choice containing the selected option, probabilities for alternatives, and confidence. A routing app can receive a Department enumeration directly, avoiding a separate step that parses generated prose into a department name.

The AnyDecisionModel README says each session holds one immutable text or JSON state. Questions are independent: an urgency answer does not become input to a later routing question. If your workflow requires that dependency, represent it explicitly in the application rather than assuming conversational memory.

Next-token backends use answer labels that limit choices to 26 options and scores to 10 levels. Trained decision heads do not have those label limits. More consequentially, local confidence describes how concentrated the probability distribution is; it is not the probability that the answer is correct. The documentation recommends fitting calibration separately from evaluation. A high confidence value alone should not authorize an automatic routing rule.

The announcement lists Swift 6.2 or later and macOS 26 or iOS 26 or later as baseline requirements. Additional requirements depend on the backend: AnyDecisionModel’s MLX build tasks need Xcode’s Swift 6.4 or later and the Metal toolchain, while Core AI needs Xcode 27 and an OS 27 runtime.

It also documents a dependency conflict: AnyDecisionModel’s CoreAI trait cannot currently be combined with AnySpeechRecognizer because they require incompatible swift-huggingface versions. Resolve the intended package combination before designing a combined speech-and-routing app around it.

Model downloads also deserve an early check. The announcement estimates about 600 MB per Parakeet model, another 590 MB for its live encoder, 0.9–2.4 GB for Qwen, and about 250 MB for Sortformer. Those are published download estimates, not measurements of installation time or device performance.

Our proposed pilot starts with recorded audio and a small set of manually labeled routing examples. Include quiet speech, overlapping speakers, relevant languages, and tickets that plausibly belong to more than one department. Keep the same inputs when comparing engines.

Measure transcription errors and speaker attribution separately. For routing, record the chosen department, alternative probabilities, and actual correctness. Fit any calibration on a separate set, then evaluate untouched examples before selecting a threshold for human review. Our guide to designing a reliable AI workflow explains how to make acceptance criteria and fallback behavior explicit.

Preserve failures as regression cases using the approach in our guide to testing AI-generated code with regression tests. Compare routing from a corrected transcript with routing from model-produced text to reveal errors introduced upstream.

Start on a supported Mac configuration, then repeat the checks on the actual target device. An iPhone release needs its own evidence. These packages merit a bounded prototype when engine comparison is a real requirement; production adoption should follow demonstrated accuracy, compatible dependencies, and acceptable behavior on your own audio.