Arc: a ~70 MB streaming speech model that transcribes links names and metadata to your own list as they're spoken, with no second model [p]
Arc-Transcriber is a ~70 MB streaming speech model that performs ASR-NL in a single step. Every 100 ms, it returns a transcript along with linked names, numbers, speakers, sound events, and whether speech was addressed to an assistant. It includes inbuilt tags for dialog act and speaker segmentation, with over 200 trained decision heads. The model is used on-device in a Mac app, where it can prime voice searches over 61k books.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年10月29日 22:51 UTC
收录当时偏移：UTC+02026年10月29日 23:01 UTC
- 发布
- 2026年10月29日 22:51
- 收录
- 2026年10月29日 23:01
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
I built Arc-Transcriber, which does ASR-NL in 1-step. Every 100 ms, one model returns the transcript plus:
- Names linked to a list you give it (a roster, contacts, product names); new names are flagged "not on the list" - Numbers with their values (scores, clocks, money, dates) - Speakers and roles; - Sound and channel events (telephone audio, clipping, poor audio) - Whether the speech was addressed to an assistant. - Inbuilt tags for dialog act, speaker segmentation and more. Total 200+ trained decision heads.
The usual ASR, then NER, then entity-linking and then LLM chain is collapsed into the model's own heads, so live "fact-checking" against your list happens inside the stream, not in a second pass.
Currently it has two variants Arc mini (which uses moonshot pretrained encoder) and second Arc-v1 (which uses parakeet pretrained encoder).
Weights: huggingface.co/WhissleAI/arc-v1.1 Code and CLI: github.com/WhissleAI/arc_transcriber
We use it on-device in our Mac app. Voice search over 61k books primes Arc with matching titles while you're still talking. Here is a video to give a glimpse on how it works.