Arc-Transcriber: ~70 MB streaming speech model integrates ASR and entity linking in one step

·2026.10.30 07:55

Key point

The ~70 MB model uses 200+ trained heads on a shared encoder to output transcripts, linked names, and metadata every 100 ms.

Arc-Transcriber performs ASR-NL in a single step, eliminating the need for a separate NER or LLM chain by using 200+ trained heads on shared encoder states.
The model emits transcripts, linked names, numbers, speakers, and sound events every 100 ms, allowing live fact-checking against a provided list during the stream.
Two variants are available: Arc mini (using moonshot pretrained encoder) and Arc-v1 (using parakeet pretrained encoder), with weights hosted on Hugging Face and code on GitHub.
The system is deployed on-device in a Mac app, where voice search over 61k books primes the model with matching titles while the user is still speaking.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.