Voice Forge

The page → · screenshots, the measurements, and the downloads.

A text-to-speech tool with the synthesiser's dials on the outside, built on a Piper/VITS voice fine-tuned from about forty minutes of my own speech.

Most text-to-speech hides the machine and gives you a speed slider. That is the wrong dial. A synthesised read sounds wrong not because the timbre is off but because every beat lands at the same metronomic distance — and speech lives in the uneven ones. So this puts the timing on the outside, measures what each control actually buys per voice, and prints the number next to the dial.

a comma on the bundled voice          145 ms
a comma on a voice not shipped here   457 ms

Same speaker, three times apart. The second voice was trained for the meditation app this engine came from and stayed with it — but the gap is the point: a measurement taken from one voice does not describe another, so Measure this voice is a button rather than a printed constant, and the app will not print one voice's figure under another's name.

The text decides the delivery, and there is no switch for it. A sentence is read rather than recited: the point of each phrase is held a little and lifted, a phrase settles at its clause mark, unstressed words give way, and a word already said in the paragraph steps back the second time. That happens per sound — a vowel can be held, a stop cannot, and holding one anyway is what turns a stretched word into a stutter.

It is the timing the model is told, not a filter over what it produced. The bundled voice's graph carries an extra input for per-token duration, and feeding it a vector of ones reproduces the stock model bit for bit, which is how the change was shown to be inert before any of it was switched on. A voice installed from elsewhere has no such input and renders exactly as it always did.

Inside a sentence, *focused words* receive their own stress and room, while [[beat:short]], [[beat]] and [[beat:long]] add authored breaths to the same model call. Pronunciations can be edited as IPA or seeded from an ordinary “sounds like” respelling and inspected before use.

There is no emotion selector. One was built — Happy, Playful, Intimate, Flirty, Angry — and it is gone from the interface because it did not earn its place. The honest reason is in docs/expression-development.md: this voice is not style-conditioned, so an expression could only reach pace, two noise scales and some EQ, and none of the three survived. The pace moved 2% of a 13% request because the graph rounds every token up to a whole frame. The noise scales only redrew the dice. The EQ read as loudness rather than character. Rebuilding it on the timing layer made the six measurably different — and still not convincing enough to keep a control for.

For controlled listening comparisons, run npm run audition from cross-platform. It creates a local HTML page with playable WAVs and a JSON measurement manifest. EQ comparisons reuse the exact same recording, with loudness matched using gain only. Separate phrase-direction comparisons disable both Piper noise scales and verify a repeatable baseline. These are experiments for listening evaluation; passing signal checks does not establish naturalness or emotional expression.

Two implementations

macos/	Swift, SwiftUI, ONNX Runtime	native Mac app
cross-platform/	TypeScript, Electron, ONNX Runtime	Windows and Linux
linux/	GTK4 + libadwaita over the Swift core in macos/	Linux, experimental

The Electron build draws its interface in HTML so it looks the same everywhere. The plan has since changed to one Swift core with a native window per platform, and linux/ is the first of those windows. So the Electron build takes fixes and parity work only. No new feature lands in TypeScript. It is retired once the GTK window passes the same fixtures and runs on the Steam Deck and Mint.

The two cores are held to each other. 20 parity checks compare phonemes byte-for-byte and durations to the millisecond — 1.637 s against 1.637 s — with the duration predictor made deterministic so the comparison means something. If they ever disagree, one of the two suites goes red.

Installing

# macOS and Linux
curl -fsSL https://raw.githubusercontent.com/snepssen/voice-forge/main/install.sh | sh

# Windows, in PowerShell
irm https://raw.githubusercontent.com/snepssen/voice-forge/main/install.ps1 | iex

Read them first if you like — they are short, and piping a remote script into a shell is exactly the thing worth being suspicious of:

curl -fsSL https://raw.githubusercontent.com/snepssen/voice-forge/main/install.sh | less

Each asks GitHub for the latest release, downloads the single archive for that machine, checks its SHA256 against the checksums published in the same release, and unpacks it. No sudo. Nothing written outside the paths it prints.

Or take a build from Releases by hand.

Two things worth knowing before you run it. On macOS the app is signed ad-hoc rather than notarised, so Gatekeeper would refuse it — the installer clears the download quarantine flag, which is what you would do by right-clicking Open, and says that it did. On Windows the installer is unsigned and SmartScreen will warn you. Code signing costs money this project does not spend.

Building

cd macos          && ./build.sh        # checks, then Voice Forge.app
cd cross-platform && npm ci && npm run check && npm start

npm run check never loads a model, which is what keeps it the cheap gate. npm run parity and npm run smoke do. npm run netcheck proves the app makes no network connections, and it is itself proved: a planted request has to make it fail before a passing result means anything.

Training your own voice

One voice ships and it is not privileged — drop a Piper .onnx and its .onnx.json in the voices folder and it appears in the picker; install one named the same as the bundled voice and yours wins. TRAINING.md is the guide, and carries this project's own measurements rather than generic advice.

The wider workshop

• Gateway Forge — guided-session authoring using the same measured speech engine.
• Voice Forge — this project.
• Protoke — lyric and narration video with a word-timed vector face.
• tools-core — corpus, audio, and speech-research utilities.

Licence

GPL-3.0-or-later. Not a preference: the app bundles espeak-ng and piper-phonemize, both GPL-3.0-or-later, so anything distributed with them has to be compatible.

The page

docs/ is the project page — one HTML file, four screenshots, no build step. Turn on GitHub Pages with the source set to main / docs and it is live at https://<username>.github.io/voice-forge/.

The three download links point at releases/latest/download/..., which always resolves to the newest release, so the page never needs editing when a version ships.

Contact

There is no updater and no crash reporter, which is deliberate — and it means a failure has no way of reaching me on its own. @snepssen on Telegram, or [email protected].