• @airesum

    Hugging Face introduces two Swift packages for speech recognition and decisions

    AI

    Published

    Hugging Face has detailed AnySpeechRecognizer and AnyDecisionModel, Swift libraries that let Apple-platform apps choose which engines handle transcription and classification-style questions.

    Hugging Face has outlined AnySpeechRecognizer and AnyDecisionModel, two Swift packages that extend its AnyLanguageModel approach to speech and decisions, in a Hugging Face blog post.

    AnySpeechRecognizer keeps Apple's speech-to-text API but lets apps pick engines such as NVIDIA's Parakeet or Alibaba's Qwen3-ASR on Core ML, and adds diarization for up to four speakers.

    AnyDecisionModel handles yes-or-no, multiple-choice and rating questions, returning typed results based on model probabilities instead of generated text.

    In tests on one M4 Max Mac Studio, Parakeet transcribed an hour of file audio in about 20 to 40 seconds; iPhones are untested so far.

    Why it matters

    Hugging Face says purpose-built models can be better, faster and cheaper than one language model doing both jobs.

    Claims & sources
    1. c4

      Hugging Face says purpose-built models for speech recognition and for decisions can be better, faster and cheaper than handing both jobs to a language model.

    2. c5

      With AnyLanguageModel and swift-huggingface both at 1.x versions, Hugging Face says it is bringing the same pattern of familiar Swift interfaces plus model choice to speech and decision tasks.

    3. c6

      AnySpeechRecognizer is an implementation of Apple's new speech-to-text API that lets developers pick the engine, for example NVIDIA's Parakeet or Alibaba's Qwen3-ASR running locally on Core ML, and it adds speaker diarization as well.

    4. c7

      For yes-or-no, multiple-choice and rating-scale questions, AnyDecisionModel gives typed answers with probabilities, derived from model probabilities rather than generated text, working with on-device models via MLX, Core AI or llama.cpp, or with decision models from Jev or Ollama.

    5. c23

      The speaker diarizer can follow up to four speakers, overlapping speech included, and a SpeakerTimeline component links each timed word to a speaker.

    6. c25

      Hugging Face took its measurements with the hear command-line tool included in the package, using an M4 Max Mac Studio running macOS 27.0.1.

    7. c26

      According to Hugging Face, Parakeet processes one hour of audio in about 20 to 40 seconds when reading a file and about 1.5 to 2.5 minutes in live mode, while Qwen3-ASR is much slower.

    8. c27

      Hugging Face cautions that the figures reflect one machine and a few datasets, and that its testing has so far been limited to macOS 27, with no iPhone tests yet.

    Original sources

    1. Hugging Face
    About this AI summary

    • AI-generated original factual summary by AI Resum, based on the linked source.

    • No review, permission, license or endorsement by the source publisher is claimed. Source content and trademarks remain with their respective owners.

    Swipe up for next1 / 1