The installer refused folders whose parent is user-owned (Desktop, Documents) because a user could swap
the whole install folder and get the elevated/audiodg-run parts replaced. Instead of refusing, the parts
that run with admin rights or inside audiodg now live in {commonpf}\VoiceChanger, which only admins can
change; the chosen folder keeps the app, Python, engine, models and data.
- VoiceChanger.iss: SysDir = {commonpf}\VoiceChanger holds apo\ (VoiceChangerAPO.dll), scripts\mic-apo.ps1
and the uninstaller (UninstallFilesDir). CheckParent removed. Upgrade detection also matches the previous
install dir from the uninstall registry (PrevAppDir) since uninstall\ is no longer inside the app folder.
Stop the audio service when either the new or the v1.0.1-layout APO exists. Ask before installing into
a OneDrive-synced folder.
- secure-dir.ps1: drop CheckParent; Prepare takes -Upgrade.
- mic-apo.ps1: -From <previous app folder> migrates a v1.0.1 apo\ or 21b4278 config\ backup (validated as
before) into the Program Files apo\, removing the emptied old apo\.
- app: admin_root() finds Program Files\VoiceChanger (falls back to the app folder for the zip layout);
hooked-mic list, mic-apo log and the reconnect button use it.
Windows 11 VM (Inno 6.7.4 build, AppVersion 1.0.2): v1.0.1 at C:\VoiceChanger -> upgrade: backup moved
byte-equal, APO re-pointed to Program Files, old apo\ scripts\mic-apo.ps1 uninstall\ gone; fresh install to
C:\Users\ci\Desktop\VoiceChanger (owner WIN-CI\ci) with full engine download: exit 0, audiodg loads
C:\Program Files\VoiceChanger\apo\VoiceChangerAPO.dll, 440 Hz APO test + Communications stream PASS;
basic-user token can rename the Desktop folder but every Program Files write/rename is ERROR_ACCESS_DENIED;
reinstall over the same folder keeps user files; uninstall restores FxProperties (only the endpoint
timestamp properties differ, same as v1.0.1) and removes both folders.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
consent-recorder: operators (ADMIN_DISCORD_IDS) create an app key on the
site; the app calls GET/POST /api/guilds with it to turn servers on/off.
Turned-off servers are left immediately. recordingNow() now checks the
user picked that server, so someone who only chose server A is not shown
as recorded while sitting in a B channel the bot records for others.
voice-changer: new "녹음 서버" dialog (record_servers.py), network in a
worker thread, key saved in config/record-servers.json; upgrade cleanup
list knows the new file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The site lists only servers the bot and the user share (single-member REST
fetch, no privileged intent). Consent requires at least one server; the
choice can be changed later and removed servers stop recording at once.
Consent text mentions the server choice, so CONSENT_VERSION is bumped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
secure-dir.ps1 Prepare used to delete every root file (and whole scripts/app/launcher/uninstall
dirs) when upgrading. It now removes only the named files of earlier distributions (plus the
uninstall logs, which run elevated) and empty dirs; any other file stays and is only re-owned
by Administrators with inherited permissions.
Windows 11 VM: 21b4278 install + root sample.wav, models\keep.wav, scripts\note.txt ->
upgrade -> SHA256 unchanged for all three (also after silent uninstall).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- installer/secure-dir.ps1 (run elevated from tmp):
CheckParent: refuse folders whose nearest existing parent is not a drive root or owned by
Administrators/SYSTEM/TrustedInstaller.
Prepare (before files are copied): owner Administrators + protected DACL (SY/BA full, Users RX),
verify, abort setup on failure; refuse links pointing outside the folder; on upgrade delete the
old executable parts (scripts/app/launcher/uninstall/root files) that 21b4278 left user-writable.
Finish: Users modify only on data dirs; verify every admin-run file is admin-owned and not
user-writable, otherwise skip APO registration and engine download.
- mic-apo.ps1: migrate config\mic-apo-backup.json to apo\ with value validation (CLSID/mode GUID/
DWORD only, own CLSID stripped), never record our own APO as the original value, upgrade the
DEFAULT-only mode list written by 21b4278 to DEFAULT/COMMUNICATIONS/SPEECH/MEDIA, remove the
old config\mic-apo.json.
Windows 11 VM: user-precreated folder -> owner BA, basic-user WRITE_DAC/WRITE_OWNER/grant/icacls/
write/rename all ERROR_ACCESS_DENIED; user-owned parent refused (CheckParent exit 20);
21b4278 install -> upgrade: backup migrated, modes refreshed, Communications capture 440 Hz,
uninstall restores FxProperties to the pre-install state.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- installer: lock the install folder (admins/SYSTEM full, Users read) and grant Users modify only on
data dirs (models/training/datasets/config/cache/env/engine/tools); download the engine as the
original (non-elevated) user; refuse non-empty folders; uninstall deletes only app-managed dirs
- mic-apo.ps1: backup/list/log moved to admin-only apo\; register SFX for DEFAULT, COMMUNICATIONS,
SPEECH and MEDIA when the endpoint has no mode list; admin TEMP for Add-Type; run log + trap
- app: '마이크 다시 연결' launches one elevated PowerShell via ShellExecuteExW(runas) with
double-quoted -File args and waits for its exit code; shows apo\mic-apo.log on failure
- tests: reconnect argument splitting (Korean/space/quote/drive root)
Verified on the Windows 11 VM: basic-user token gets ERROR_ACCESS_DENIED on scripts/apo/exe/
uninstall/app and write access only on data dirs; Communications-category capture receives the
converted signal (440 Hz); reconnect exits 0 from a Korean+space install path.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- apo/: VoiceChangerAPO.dll (SFX APO on the real mic, built with mingw). When the app converts,
the mic's own stream carries the converted voice; otherwise it passes the real voice through.
Shared-memory protocol in vccore.h, Python side app/mic_bridge.py
- scripts/mic-apo.ps1: register/unregister on every capture endpoint with admin value rights only
(no ACL changes unless FxProperties is missing), originals backed up and restored
- app: mic is the conversion target (WASAPI names), no output device, '마이크 다시 연결' button,
44.1k<->48k streaming resample, caches/temp redirected into <install>/cache
- installer/VoiceChanger.iss (Inno Setup): choose install folder, register APO, download engine
into that folder, uninstall restores mic settings and asks before deleting models
- install.ps1: uv python --no-bin --no-registry, temp/caches inside the folder, uv cache clean
- tests: C core compiled with gcc against the Python bridge; converter pump 44.1k round trip
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Applio's realtime engine treats incoming audio as 48kHz; opening a 44.1kHz
device at its default rate skewed pitch and timing. Streams now always open
at 48kHz (MME/DirectSound resample, WASAPI auto_convert), with a clear error
if a device refuses it.
- KSS/Zeroth example datasets write a completion marker (with minutes) last;
a partial folder from an interrupted run is wiped and re-extracted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /delete discards the in-progress utterance before deleting files
- keep the current channel even when the bot itself fills it (joinable=false)
- bind OAuth state to the browser via a short-lived cookie
- reject consent submitted from a page showing an older consent version
- regression tests for all four
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Discord OAuth login -> explicit consent (3 checkboxes, versioned text) ->
bot follows consenting users into voice channels and records only their
per-user streams. Revoke stops recording immediately and optionally deletes
files; consent log kept; export to training wav only for active consents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Remove the 'instruct' style field: Qwen3-TTS Base generate_voice_clone
silently ignores it, so the UI option did nothing.
- Serialize generations with a lock so concurrent requests queue instead
of doubling RAM/VRAM on one shared model.
- Validate empty text/ref_text ourselves so the API returns the Korean
400 messages instead of FastAPI's generic 422.
- UI: explain that mic recording needs https/localhost instead of a raw
TypeError on plain-http LAN access.
- README: document mic/https limit, queueing, and measured CPU speed.
Verified on CPU (Python 3.12, torch 2.14 CPU, 0.6B Base): wav and mp3
references, language ko/auto, two concurrent requests -> all 200 and
Whisper transcripts match the requested text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Extract the voice-cloning feature from jamiepine/voicebox (Qwen3-TTS Base
engine) into a small standalone app: FastAPI backend (engine.py + app.py)
wrapping create_voice_clone_prompt/generate_voice_clone, and a single-page
UI (upload or mic-record a reference clip + transcript -> synthesize any
text in that voice). Supports 10 languages incl. Korean; model loads lazily
and downloads on first use.