Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms
Source
huggingface.co
Date
Why it matters
Apple-platform developers can run speech-to-text and typed yes/no or multiple-choice decisions on-device through Swift APIs. This avoids routing both jobs through a generative LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition →.
Key takeaways · AI-distilled
In Hugging Face's tests on an M4 Max, English-only Parakeet Unified had a 1.43% word error rate on clean LibriSpeech files versus 2.12% for Apple's transcriber. On AMI meeting recordings with overlapping speech, every engine tested exceeded 23%. iPhone was not yet tested.
Per Hugging Face, Parakeet transcribes an hour of audio from a file in about 20 to 40 seconds, using about 100 MB of memory once Core ML compiles the model. An optional 8-second live-input encoder cuts live analysis time by about a third at a somewhat higher word error rate.
Qwen3-ASR covers languages Parakeet does not, such as Mandarin, Cantonese, Japanese and Korean, but returns one final result per segment of up to 40 seconds with no word timing or confidence, runs on GPU, and is much slower, so Hugging Face suggests it for files over live input.
AnyDecisionModel's next-tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → backends label each allowed answer with a single token (yes/no, A to Z, or 0 to 9) and read its probability, which caps choices at 26 options and scores at 10 levels. Models with a trained decision head, like Cloudflare's Clef, have no such limit.
To counter position bias, rotationDebiasing asks a choice question once per rotation of the options and averages the results, costing one prompt evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → per rotation. The confidence value is entropy-based: it shows how concentrated probabilities are, not correctness.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
