Proposal: URML (substrate-neutral robot intent) manifest declaration for Whisper as the multilingual listen substrate

[idoco2003]

Hi @openai/whisper team,

Proposing a URML v0.1 capability-manifest mapping for Whisper over openai/whisper. URML (Apache-2.0) is a substrate-neutral spec for robot intent: a typed primitive vocabulary plus a Layer-1 capability manifest and a validator that gates programs against the manifest before any actuator publishes.

Whisper is the reference multilingual STT. URML's Layer-2 listen primitive consumes Whisper transcripts as the input to URML's natural-language bridge; URML's Layer-4 reserves multilingual slots (English content; Hebrew, Spanish, Japanese, Mandarin reserved in v0.1) that Whisper's 99-language coverage maps onto directly. Engaging through Discussions since Issues are disabled on this repo.

This is proposal-only, the first RFC of URML's Move #12 outreach (16 RFCs covering speech / translation / robot-command-library substrates for URML's NL layer).

Full RFC with manifest mapping, three alternatives, and the inference-runtime fragmentation discussion: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0153-whisper-outreach.md

Questions worth maintainer input on:

1. Engagement channel. Discussions Q&A or Show-and-tell? URML's post is a design-discussion request, not a bug report.
2. STT-engine-class declaration shape. URML needs to distinguish the Whisper family from the inference runtime (openai-reference vs. CTranslate2 vs. ggml). Does the OpenAI team have a preferred convention?
3. Multilingual labelling. Whisper auto-detects source language; URML's manifest declares an explicit stt_languages list for static validation. Is the explicit list a useful downstream signal, or is auto-detect the canonical default?
4. Decode-mode boundary. Whisper's built-in translate mode overlaps URML's separate translation-engine layer. Is one of these modes the canonical URML default, or should the manifest support both?
5. Cadence expectation. Is openai/whisper actively monitoring Discussions, or has the active community moved to faster-whisper / whisper.cpp?
6. Anything else.

Ido Yahalomi (URML maintainer, urml.dev, greenvh@gmail.com)

---

AI-assisted prose, maintainer-reviewed before posting (see VIBE.md). Human-only correspondence available on request.