01 / CARTOGRAPHY
Gateway Forge
A guided-meditation assembly system and journal that keeps the published map beside what was actually found—without forcing them into one answer.
Independent software workshop
Voices, signals & strange maps.
I build local-first tools that put the interesting machinery on the outside—so it can be measured, questioned, and changed.
Eight public projects
The workbench
Each project has its own repository and field notes. The apps stay local; the measurements, trade-offs, and failures are left visible.
02 / SPEECH
Voice Forge
A text-to-speech instrument with the synthesiser’s timing controls exposed and measured instead of hidden behind one generic speed slider.
03 / FACE
Protoke
proto + karaoke. A karaoke and narration renderer fronted by a protogen visor face — two eyes and a mouth drawn as vector outlines and burned in by the same libass pass that draws the lyrics, so there is no image sequence to composite.
◤‿◥ visor^w^ happy▶_◀ sharp
¬‿¬ smugo_o wide-_- sleepy
owo curious^o^ delighted^_^ content
¬‿¬ smugo_o wide-_- sleepy
owo curious^o^ delighted^_^ content
04 / DELIVERY
Media Preflight
Check a finished file against where it is going — YouTube, ACX, EBU R 128, or a standard of your own — and get the exact timestamps of what is wrong. Then, optionally, a corrected copy that is measured again from scratch before it claims to be fixed. The original is never touched.
✓ -23.1 LUFS✓ -1.4 dBTP⚠ 2.1 s black
✕ 48 kHz✓ h264⚠ drift 0.4 s
✓ fast start✕ mono✓ no clipping
✕ 48 kHz✓ h264⚠ drift 0.4 s
✓ fast start✕ mono✓ no clipping
05 / INTAKE
siphon
Paste a link and get the file — or point it at a folder you already have and get the same thing in a different shape. It is one pipeline either way, and it does not re-encode what it does not have to: asking for an mp4 that is already H.264 and AAC moves the streams into a new container and changes nothing else. Playlists from Deezer, Spotify and Tidal arrive tagged, covered and foldered, and a track it is not sure about stops and asks before it fetches.
✓ audio copied✓ video copiedalready h264
✓ mp4 ⟶ mp4nothing re-encoded16 s
⚠ matched 69%worth a listen✓ cover embedded
✓ mp4 ⟶ mp4nothing re-encoded16 s
⚠ matched 69%worth a listen✓ cover embedded
06 / FILING
auto-sort
Empties a Downloads folder and keeps it empty. It works out what each file is from its bytes rather than its name, reads what documents call themselves, and once three of them share a heading they get a folder of their own. When it learns something new it goes back over what it already filed. Every move is written down and can be undone, and nothing is ever deleted: what goes, goes to the bin.
✓ scan.pdf⟶ Scans/2026read, not named
✓ 3 invoices⟶ new folderlearnt
⚠ encryptedset aside↶ undo any move
✓ 3 invoices⟶ new folderlearnt
⚠ encryptedset aside↶ undo any move
07 / INTERPRETING
Interpreter
Live interpreting into English that waits for a complete thought, the way a human interpreter does, instead of translating every half-sentence. Polish, Hungarian or any of 25 languages around the table arrive in your ear a few seconds behind; English is left alone. Recognition, translation and the voice all run on the iPhone or Mac in front of you.
✓ halb elf⟶ 10:304.2 s behind
✓ 916 Englishleft quietcontext kept
⚠ 7-5please repeatnot spoken
✓ 916 Englishleft quietcontext kept
⚠ 7-5please repeatnot spoken
08 / FIELD TOOLS
tools-core
The useful pieces that fell out of the larger work: caption checks, media repackaging, audio tagging, file identification, duplicate finding, asset grouping and tray icons, alongside audio forensics, corpus preparation and voice research.
tools-core/
├── audio-analysis/
├── audio-tag/
├── bundle-list/
├── caption-check/
├── exact-duplicates/
├── file-identify/
├── gateway-authoring/
├── piper-workflows/
├── remux/
├── site-capture/
├── site-fonts/
├── tray-icon/
└── tts-research/
├── audio-analysis/
├── audio-tag/
├── bundle-list/
├── caption-check/
├── exact-duplicates/
├── file-identify/
├── gateway-authoring/
├── piper-workflows/
├── remux/
├── site-capture/
├── site-fonts/
├── tray-icon/
└── tts-research/
tools-core
Open the toolbox
These are working instruments, not a framework. Take the one that answers your question, give it explicit paths, and keep the generated material out of Git.
git clone https://github.com/snepssen/tools-core.git && cd tools-core
Caption Check
Find subtitle cues that are too fast, too long, overlap in the same position or run past the end. Check SRT, VTT and ASS locally, with the exact times and configurable thresholds.
python3 caption-check/caption-check.py subtitles.srtRemux
Put a recording in another container without re-encoding. Every supported track is copied; an incompatible track stops the operation. Preview the decision, then write a new file.
python3 remux/remux.py recording.mp4 recording.mkv --planExact Duplicates
Find files with identical bytes across folders. See every copy and its logical size, with hard links counted once. Nothing is deleted, moved or chosen for you.
python3 exact-duplicates/exact-duplicates.py ~/DownloadsBundle List
See which files belong together: a film and its subtitles, a RAW/JPEG pair, a model and its sidecars, or a sequence of frames. Read the groups or export a JSON listing.
python3 bundle-list/bundle-list.py ./assets --groups-onlyAudio Tag
Stop “Unknown artist”. Title, artist, album, year and track number for untagged WAV, MP3 and FLAC files, read from the folder layout, the file names and the order they were made. Preview first, then write in place.
python3 audio-tag/audio-tag.py ~/Music --artist "Your Name"Beat analyser
Find persistent stereo tone pairs, calculate their differential, and separate likely entrainment signals from drifting content.
python3 audio-analysis/analyse-beat.py track.flac --jsonCorpus aligner
Use authored text as truth, ASR only for timing, and quiet gaps for safe cuts—then preserve a manifest of every decision.
python3 audio-analysis/align-corpus.py read.wav words.json script.md outputPiper workflows
Train, resume, export, render, and A/B checkpoints while retaining fixes for checkpoint loading, dynamic shapes, and telemetry.
bash piper-workflows/check-in.sh RUN CONFIG OUTTruth probes
Dump real MLX tensors, tokenizer slices, and speaking pace so a native port can be compared stage by stage instead of trusted by ear.
python3 tts-research/qwen3-tokenizer-truth.py --gateway-root PATHLibrary helpers
Index evidence, convert measured signals into playable profiles, verify corpus labels, and safely preview provisional briefings.
python3 gateway-authoring/seed-briefings.py PATH --dry-runFile identifier
Say what a file is from its bytes, not its name: the real format when the extension lies, the heading a document gives itself, when it happened and where it came from. auto-sort’s own identifier, unchanged, and it never writes.
python3 file-identify/identify-file.py FILE --factsTray icon
Put an icon with a menu in the macOS menu bar, the Windows notification area or a Linux panel, each entry running a command of yours. Nothing installed: auto-sort’s own tray, speaking to the Objective-C runtime, Win32 and D-Bus with the standard library.
python3 tray-icon/tray-icon.py "Backups" --item "Back up now" "~/bin/backup.sh"Small live specimens
Try the ideas
Two deliberately tiny sketches from the larger projects. Nothing is uploaded and nothing follows you away from this page.
Two tones, one perceived beat
Pick a differential, put on headphones, then start the quiet signal. The left and right channels differ by the selected amount.
ready · carrier 100 Hz · differential 4.0 Hz
Headphones required to perceive the differential. Starts very quietly, stops automatically after 30 seconds, and stops when the page is hidden.
A lyric line is a clock
Play the timing sketch. The words are fixed; only their measured windows decide what is active.
Build a little signal,
then give it room to breathe.
then give it room to breathe.
00:00.000 / 00:05.200
The face is a subtitle track
No image sequence and no video layer: two eyes and a mouth, drawn from a dozen numbers and burned in by the same libass that draws the lyrics. These are Protoke's own frames, played back.
64 frames · 3 shapes · one loop
A recurring method
How I build
The subjects change. The rules underneath them have stayed remarkably consistent.
Keep it local
Personal voice, private journals, and unfinished material stay on the machine unless there is a clear reason for them to leave.
Measure the claim
A waveform, duration, tensor, or failing check is more useful than “sounds about right.” Unknown stays a named state.
Show the mechanism
Good tools expose the meaningful controls and preserve the source material, instead of hiding both behind a convenient button.
The wider workshop
Nine projects, five foundations, one local-first practice.
Move between guided-meditation authoring, measured speech, live interpreting, a voice companion, lyric video, delivery validation, keeping a Downloads folder empty, and the small field tools that support all of them, on one voice, one speech engine and one hearing layer they share.
01 / JOURNEYSmacOS
Gateway Forge
Assemble guided sessions and keep published maps beside what was actually found.
Explore Gateway Forge →02 / SPEECHmacOS · Linux · untested: Windows
Voice Forge
A transparent Piper workbench for timing, pronunciation, and honest audio export.
Explore Voice Forge →03 / VIDEOmacOS · tested on Linux, Windows
Protoke
Karaoke and narration video with a vector visor face that mouths every word.
Explore Protoke →04 / DELIVERYmacOS · tested on Linux, Windows
Media Preflight
Check a finished file against where it is going, and correct it without touching the original.
Explore Media Preflight →05 / INTAKEmacOS · tested on Linux · untested: Windows
siphon
Fetch from a link or convert what you already have, without re-encoding what it does not have to.
Explore siphon →06 / FILINGmacOS · Linux · tested on Windows
auto-sort
Empty the Downloads funnel into structure the folder taught it, and undo any of it.
Explore auto-sort →07 / INTERPRETINGmacOS · iPhone · iPad · untested: Android
Interpreter
Hear the table in English a few seconds behind, cut at the end of each thought, on your own phone.
Explore Interpreter →08 / COMPANIONmacOS · iPhone · iPad
companion-forge
A local voice companion for Mac, iPhone and iPad, with a protogen LED-visor face.
Explore companion-forge →09 / FIELD TOOLSPython · shell
tools-core
Small utilities for audio inspection, corpus work, authoring, and repeatable checks.
Open the toolbox →Foundations
The parts the projects are built from. They have no page of their own: each is a repository, and the projects are where you see them at work.
- voice-polyglotGitHub ↗ Teaching one small voice to speak eight languages, and releasing it for every app that speaks. Used by Interpreter and companion-forge
- voice-coreGitHub ↗ The speech engine the apps share: espeak-ng, the phonemizer and Piper inference, from one source. Used by Gateway Forge, Voice Forge, Interpreter and companion-forge
- speech-coreGitHub ↗ The hearing layer Interpreter and companion-forge share: ObjCExceptions, SileroVAD, SampleBuffer and ModelStore, from one source. Used by Interpreter and companion-forge
- workshop-coreGitHub ↗ How the Python tools find the programs they need, and offer to install what is missing, from one source. Used by Protoke, Media Preflight and siphon
- site-coreGitHub ↗ The builder and project network every page here shares, kept in one place. Used by every page in the workshop
Get in touch
If something looks wrong, say so
Bug reports belong in the project that produced them. For everything else, email or Telegram reaches me directly.