Hugging Face shipped Swift packages that let Apple apps swap in an open speech-recognition engine and read decisions from model probabilities instead of generated text
AspireAI community · Hugging Face
Hugging Face announced AnySpeechRecognizer and AnyDecisionModel on October 29, extending its Swift tooling from language models into two tasks that language models are often asked to do suboptimally. AnySpeechRecognizer mirrors Apple's SpeechAnalyzer and SpeechTranscriber APIs while letting the app choose the engine, including NVIDIA's Parakeet TDT 0.6B v3 and Parakeet Unified, Moondream's Parakeet Ultra, and Alibaba's Qwen3-ASR at 0.6B and 1.7B, with speaker diarization through Sortformer added on top; the Core ML conversions of all of them are published on the Hub with source checkpoints and checksums. AnyDecisionModel answers yes-or-no, multiple-choice and rating-scale questions with typed results and probabilities derived from the model's output distribution rather than from generated text, running on device through MLX, Core AI, llama.cpp, Jev or Ollama. The downloads are not small, roughly 600 MB per Parakeet plus about 590 MB for the live encoder and 0.9 to 2.4 GB for Qwen3-ASR, and the speech engines differ in ways the app has to absorb, since Qwen3-ASR returns one final result per segment of up to 40 seconds with no word timing or confidence while Parakeet streams volatile results and timings.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.