TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Hugging Face report introduces AnySpeechRecognizer and AnyDecisionModel, Swift packages that let developers select among model engines for speech recognition and structured decisions on Apple platforms. The packages support on-device models, but available features, speed, language coverage, model size and licensing differ by engine.
Hugging Face has announced AnySpeechRecognizer and AnyDecisionModel, two Swift packages that give Apple-platform developers a shared interface for selecting different speech-recognition and decision-model engines. The packages support on-device models, including Core ML speech models and decision models running through MLX, Core AI or llama.cpp, offering an alternative to relying on a single model provider for these tasks.
AnySpeechRecognizer mirrors Apple’s newer SpeechAnalyzer API, including components for transcription, dictation, speech detection and asset installation. Apple’s Speech framework remains the default engine, but developers can substitute other supported models while keeping much of the surrounding workflow—installing model assets, processing audio and reading results—consistent. The package’s stated aim is to make changing engines require relatively few code changes.
Supported speech options include NVIDIA’s Parakeet family and Alibaba’s Qwen3-ASR, converted to Core ML and published on the Hugging Face Hub. Parakeet v3 recognizes 25 European languages and detects the language automatically; Qwen3-ASR is offered in 0.6-billion- and 1.7-billion-parameter versions and supports 30 languages and 22 Chinese dialects, according to the report. The capabilities differ: Parakeet supports live and file input, word timing and confidence information, while Qwen produces a final result for audio segments of up to 40 seconds and is described as slower, making it more suited to file transcription.
The second package, AnyDecisionModel, is designed for yes-or-no, multiple-choice and rating-scale questions. Rather than generating a free-form text response and interpreting it afterward, the package derives typed answers from model probabilities, the report says. It supports on-device execution through MLX, Core AI or llama.cpp, as well as decision models from Jev or Ollama. The source does not provide a full list of supported models, comparative accuracy results or performance benchmarks for this package.
More Choice for On-Device Speech Apps
The packages address a practical development problem: speech recognition and structured decisions are often treated as language-model tasks, but applications may benefit from purpose-built systems. A support app, for example, could transcribe a caller and then route the call using a defined set of choices. Separate model options let developers select engines suited to each job instead of routing every step through one general-purpose language model.
For Apple developers, a shared Swift interface could reduce the work of testing or changing providers. On-device processing can also keep inference on a user’s hardware rather than sending every audio segment or decision request to a hosted service. That can matter for privacy, connectivity and operating costs, though the report does not establish that every supported model will be faster, more private in every deployment, or less expensive than a hosted alternative. The trade-offs depend on hardware, model size and the chosen engine.
The announcement also makes model packaging and distribution part of the developer experience. Models are available through Hugging Face repositories, with downloads initiated when an app requests them. The stated sizes range from roughly 600 MB for each Parakeet model to 0.9–2.4 GB for Qwen3-ASR, with a separate live encoder adding about 590 MB to Parakeet downloads. Those requirements could affect storage, first-use setup and whether an app can offer a given model to all users.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
As an affiliate, we earn on qualifying purchases.
The packages extend the approach used by AnyLanguageModel, which the report describes as a Swift interface for using local and hosted language models. The authors say AnyLanguageModel and swift-huggingface have reached 1.x releases and present the new speech and decision packages as an expansion of that model-choice approach.
Apple introduced SpeechAnalyzer with macOS 26 and iOS 26, according to the report. Its framework provides transcription and detection modules and uses Apple’s own engine. AnySpeechRecognizer retains a similar interface but adds selectable recognition engines. Developers can also add another speech provider by implementing the package’s model and executor protocols, the report says.
The converted models have different licenses, including CC BY 4.0, Apache 2.0 and NVIDIA’s Open Model License. Developers need to check the terms for the specific model they use rather than assume that every model in the package has the same license. The report says model repositories identify their source checkpoints and include file checksums, and that the package pins models to fixed revisions.
“AnySpeechRecognizer implements Apple’s new speech-to-text API with a choice of recognition engine.”
— Hugging Face report
Performance and Model Trade-Offs
The announcement does not provide a complete independent benchmark comparing the speech engines, decision models or Apple’s built-in recognizer across devices. Its claims about relative accuracy and speed should be read as statements in the report, not as a broad comparative evaluation. The report says Parakeet Ultra was more accurate in most of the authors’ measurements, but does not give the full results or test conditions in the supplied material.
It is also not clear from the source how AnyDecisionModel performs across specific tasks, what models are available through each supported runtime, or how its probability-derived answers compare with other methods. Device compatibility, memory needs and real-world battery impact are not detailed. Model licenses and download sizes vary, and the source does not establish whether every listed model is suitable for every commercial use.
The supplied report text ends during its discussion of speaker diarization, so it does not provide the complete account of diarization behavior or all implementation details. The date of publication is also absent from the supplied material.
Testing Models in Swift Apps
Developers can evaluate the packages using the examples and model repositories linked from the Hugging Face report. The next practical steps are to test recognition quality with representative audio, measure latency and memory use on target Apple devices, and check whether a model’s language coverage, live-input support and licensing fit the app.
For speech applications, developers will also need to decide when model downloads occur and how to handle their storage requirements. The report says assets are downloaded only when an app requests them and stored in the standard Hugging Face cache. Further evidence on real-world performance, compatibility and the breadth of AnyDecisionModel’s model support will depend on package documentation and developer testing; the supplied source does not give a timetable for additional releases or benchmarks.
Key Questions
What are AnySpeechRecognizer and AnyDecisionModel?
AnySpeechRecognizer is a Swift package for selecting speech-recognition engines through an interface modeled on Apple’s SpeechAnalyzer API. AnyDecisionModel produces typed answers to yes-or-no, multiple-choice and rating-scale questions using model probabilities.
Can the models run on an Apple device without a hosted service?
Yes. The report says the packages support on-device options, including Core ML speech models and decision models using MLX, Core AI or llama.cpp. The available runtime and features depend on the model selected.
Which languages do the speech models support?
According to the report, Parakeet v3 recognizes 25 European languages. Qwen3-ASR supports 30 languages and 22 Chinese dialects. Language coverage varies by engine, and developers should confirm the selected model supports their use case.
Can AnySpeechRecognizer transcribe live audio?
Parakeet supports file and live input, with volatile results and word-level information. The report says Qwen3-ASR instead returns a final result for segments of up to 40 seconds and is slower, so it is better suited to file transcription than live use.
Are all the models licensed the same way?
No. The report lists different licenses, including CC BY 4.0, Apache 2.0 and NVIDIA’s Open Model License. Developers should review the license for each model repository before using it.
Source: huggingface.co
