145ms
a comma on the bundled voice
Voice Forge
it says
snˈɛpsən
in my voice
A text-to-speech tool with the synthesiser’s dials on the outside — built on a voice trained from about forty minutes of my own speech.
Why it exists
The interesting part isn’t the voice. It’s the timing.
Most text-to-speech hides the machine and gives you a speed slider. That’s 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 app puts the timing on the outside: how long a comma is worth, how much silence sits between two sentences, how much room a paragraph gets. Then it measures what those controls actually buy, per voice, and prints the number next to the dial.
Every figure on this page was measured, not estimated. Where something has not been tested, it says so.
Measured
What a comma is worth
Piper has no pause table. A comma reaches the model as a phoneme like any other and the duration predictor decides on the spot. The only honest way to answer is to render a line with the mark and the same line without, and subtract.
190ms
a colon, same voice
457ms
a comma on a voice not shipped here
Two voices from the same speaker sit three times apart at a comma, and 194 words a minute against 148 on identical copy. Only the first ships here — the second was trained for the meditation app this engine came from, and stayed with it. That gap is the entire reason Measure this voice is a button rather than a printed constant: a figure taken from one voice does not describe another, and the app will not print one under the other.
−19.99LUFS
a 1 kHz tone at −20 dBFS RMS, at 48 kHz — the loudness meter is real ITU-R BS.1770-4, not an approximation
−19.96LUFS
the same tone at 22.05 kHz — the filters are derived per rate, never copied from the published 48 kHz table
The app
Dials, and what they cost
Left: the script, and what it became — every sentence with its length, pace and peak, playable and re-rollable on its own. Right: the dials, each with the measured consequence printed underneath.
Light and dark, and the cross-platform build beside it. The interface is drawn rather than borrowed from a toolkit, so it is the same on every platform instead of three apps that happen to share a name.
You say potato
espeak gets most words right unaided — it manages Siobhan without help. What breaks is proper nouns, brand names, acronyms, and the words people simply disagree about. So the dictionary starts from the words in your script rather than an empty box, and it takes IPA, because IPA is exactly what the model consumes. What you type is what it is handed.
Left, what espeak says. Right, what you want instead. The green line is the part that earns its keep: it runs the substitution and reports where it actually landed, because stress shifts in context and an entry that quietly does nothing is the exact failure this feature exists to prevent. Entries are global by default, with per-script overrides that shadow them.
Every character is checked against the voice’s own 161 symbols before it reaches the model. The
/slashes/ you would paste from a dictionary are refused, and so is a syllable dot — to this model . is the full-stop phoneme, so it would break the sentence in half mid-word. An ASCII r only warns: it is the trilled r of Spanish, real IPA, and somebody might mean it.Three things it does that other tools don’t
Pauses, pronunciation, and an honest export
Pause control that admits its units
Silence between sentences is exact — they are separate calls to the model. A comma is not: it happens inside one call, so the app hands the model more room and measures what that buys, rather than promising milliseconds it cannot deliver.
A pronunciation dictionary in IPA
Global entries with per-script overrides. Every character is checked against the voice’s own 161 symbols, because the phoneme mapper silently drops what it doesn’t know — the slashes you’d paste from a dictionary vanish without a word.
Export that tells you before, not after
A dry voice sits near −24 LUFS with peaks around −6 dBFS, so a −14 streaming target needs gain the true-peak ceiling won’t allow. The app shows you that before you export, and says the shortfall in dB.
Downloads
Get it
One line in a terminal. It asks GitHub for the latest release, takes the single build for that machine, checks its SHA-256 against the checksums published in the same release, and unpacks it. No sudo, and nothing written outside the paths it prints.
macOS & Linux
curl -fsSL https://raw.githubusercontent.com/snepssen/voice-forge/main/install.sh | sh
Windows · PowerShell
irm https://raw.githubusercontent.com/snepssen/voice-forge/main/install.ps1 | iexPiping a remote script into a shell is exactly the thing
worth being suspicious of, so read it first — both are short and do nothing
clever: install.sh
· install.ps1.
On macOS it also clears the download quarantine flag, because the app is signed
ad-hoc rather than notarised and Gatekeeper would otherwise refuse to open it.
Or take a build by hand
Windows
Built and configured. Nobody has run it yet — if you do, tell me what happened.
| Platform | Build | Status | What was checked |
|---|---|---|---|
| macOS 15, Apple Silicon | Swift / SwiftUI | Verified | Development, all suites, packaged app, real voiceover exported |
| SteamOS 3, Steam Deck | Electron / TypeScript | Verified | Launches, renders, plays, exports — and sounds right, which is the check that matters |
| Windows 10 / 11 | Electron / TypeScript | Untested | Packaging configured, CI written, binary never executed |
The Linux pass raises the odds for Windows: it proved the parts most likely to be platform-sensitive — the ONNX native binary, the WebAssembly phonemizer, espeak data loading by path, audio out. What’s genuinely untried on Windows is path handling and the installer.
Your own voice
The app will use any Piper voice you give it
One voice ships, and it isn’t privileged: drop a Piper .onnx and its .onnx.json into the voices folder and it appears in the picker. Install one named the same as the bundled voice and yours wins.
- Record 30–60 minutes of your own clean speech, reading prepared text. Consistency beats quantity — one microphone, one room, one session.
- Cut it into clips at silence, never mid-word, and don’t trim tight. This is the step that decides whether it sounds like you, and the one most guides cover in a sentence.
- Fine-tune from a pretrained checkpoint. Measure iterations per second over one partial epoch before committing days — a wrong batch size doesn’t fail, it quietly takes a month.
- Export to ONNX, install, and measure it. The pause figures are per voice and don’t transfer.
The full guide ships inside the app and carries this project’s own measurements rather than generic advice — including the batch-size cliff that cost seventeen times the training speed for one power of two.
How it’s built
Two implementations, one set of measurements
The Mac app is Swift and stays Swift. Everywhere else is Electron and TypeScript, with the interface drawn in HTML so it looks the same rather than borrowing a platform toolkit. The engine is Piper/VITS through ONNX Runtime, with espeak-ng for phonemes.
161
checks on the Swift core
119
on the TypeScript core, deliberately the same assertions
18
parity checks holding the two engines to each other
The parity suite is the one that matters. Phonemes are compared byte-for-byte and durations to the millisecond — 1.637 s against 1.637 s, 4.888 s against 4.888 s, with the duration predictor made deterministic so the comparison means something.
One finding worth the whole exercise
The npm phonemizer bundles a newer espeak-ng whose American English rules moved the NORTH/FORCE vowel from ɔːɹ to oːɹ. Across 35 ordinary words, 8 differed — four, before, more, door, important, course, report, support — and the model was trained on the first. Shipping the Mac build’s espeak data with the WebAssembly makes it byte-identical again.
It fails open: point it at the wrong directory and nothing errors. The app runs, speaks, and is wrong in some of the commonest words in English. Eight parity checks own those words for exactly that reason.
It makes no network connections
No updater, no crash reporter, no analytics. Chromium’s background chatter is switched off by name, and the socket layer refuses any non-loopback connection outright. The check that proves it was itself proved: a deliberately planted request had to make it fail before the passing result was believed.
Get in touch
If it stops working, tell me
There’s no updater and no crash reporter, which is deliberate — and it means a failure has no way of reaching me on its own. The app has a Help sheet that copies its version, platform and voices to the clipboard for you to read before you send it. Your script and your dictionary are never in it.