Product: extract voice transcription to its own CVM (4 GB)

[daopunk]

Goal

Spin voice transcription out of the main Signal bot into a separate product with its own repo (or clearly bounded package) and its own Phala CVM at 4 GB RAM (tdx.medium).

Today Whisper runs in the same CVM as the bot + Signal CLI + translation. Long transcription jobs contend with Language Thread fan-out / NEAR translation and bog the box down.

Why

• Transcription is latency-heavy (model load + audio decode).
• Translation / Language Threads need predictable headroom for NEAR + Signal group fan-out.
• Operators should be able to run transcription alone, or pair it with either translation product without taking the whole stack.

Scope

• Carve Whisper / voice pipeline into a standalone service (own Docker image, own compose, own CVM).
• Own repo or monorepo workspace that can ship independently.
• Deploy target: Phala CVM 4 GB RAM (tdx.medium), not sharing memory with the translation bot.
• Define a clear contract for the Signal bot(s) to call transcription (HTTP/internal API, queue, or Signal-side handoff — decide in implementation).
• Document ops: register, health, attestation, resource limits.

Interoperability

This product should be composable with both translation products:

1. In-group translation (V1-style auto-translate in one chat)
2. Language Threads (sidecar / threaded translation — see related issue / feat: Language Threads — mutual-aid language sidecars #7)

Someone should be able to say: “I want transcription + in-group translation” or “transcription + Language Threads” without forcing all three into one CVM.

Out of scope (for this issue)

• Changing Language Thread relay semantics
• Merging transcription back into the translation CVM

Related

• Current monolith stack: whisper-api + signal-bot in one compose
• Language Threads PR: feat: Language Threads — mutual-aid language sidecars #7

[daopunk]

Translation products (in-group vs Language Threads): #9 · Language Threads PR: #7