vtspeak

Imported from SeasonalNet/vtspeak (AGENTS.md). Install upstream with npx skills add SeasonalNet/vtspeak. Copyright stays with the author.

Repository instructions

Scope

This repository is a reverse-engineering research and reimplementation workspace for the legacy VoiceText engine and its model data. Keep changes focused on understanding the binaries, documenting evidence, and building independent analysis tooling.

Preserve analysis inputs

• Treat binary/, data-common/, and data-paul/ as read-only vendor inputs. Do not patch, normalize, rename, repack, or overwrite their files.
• Keep generated Ghidra projects, caches, and raw exports under tools/revkit/work/; that directory is intentionally ignored.
• Put curated, reviewable explanations under docs/ and link them from the root README.
• Do not relicense or attempt to relicense vendor work (e.g. model data, voice corpus), it should currently be assumed and treated as proprietary.

Reverse-engineering evidence

• Distinguish direct observations from inferences.
• For binary layouts, record the reader function/address, widths, endianness when established, and a sample file/offset that supports the claim.
• Describe Ghidra output as pseudocode. Decompiled names, types, and control flow may be imperfect.
• Validate proposed parsers and decoder behavior against multiple files or known engine output before describing a format as fully recovered.

Repository safety

• Inspect git status before and after edits; preserve existing user work.
• Do not commit, push, tag, publish, or open a pull request unless explicitly asked.
• Do not add large vendor inputs or generated projects to Git. The root and toolkit ignore rules reflect this boundary.
• Do not weaken security controls or silence quality tools to make an analysis command succeed.