A voice trainer that keeps your voice yours.
Voice, as a practice.
Mira is an open-source, end-to-end-encrypted voice-training companion for trans people — real-time feedback on five vocal dimensions, with no score and no upload. We started building it because we wanted it to exist.
We built it for ourselves.
Most voice-training tools make one of two trades. The clinical ones score you — a number, a pass, a failure state when you miss. The convenient ones ask you to send your voice to someone else's server. For something as personal as practising your own voice, both trades are bad ones.
Mira started because someone here was doing this practice and nothing out there fit. So it is built by a person who actually uses it — shaped around what the practice needs rather than what demos well: privacy, patience, and feedback that treats you as the expert on your own voice.
“Voice is a practice, not a test.”
— the idea Mira is built around
Three tools, done well.
Mira's full scope is bigger than its first release — and that is on purpose. Phase 1 is three tools, finished, instead of ten tools half-built. Here is what ships first.
The practice tool
Phrase-based exercises with live feedback. You read a line out loud; Mira listens; five sliders move as you speak. It is the core loop — say something, see what was heard, adjust, say it again.
The library
Your practice, kept. Recordings live in a library only you can open — encrypted with a key derived from your password. It is how you hear last month against this morning, instead of guessing whether anything changed.
The listening tools
You cannot change what you cannot hear. These train the ear — pulling resonance apart from pitch, learning what weight sounds like — so the feedback starts to make sense from the inside, not just on a slider.
Five sliders. Honest math underneath.
Five sliders look simple. The signal processing under them is not. Mira analyses live microphone audio in the browser and resolves it into five dimensions — and each one is a different measurement problem.
How high the voice sits
The fundamental frequency, in hertz — the measurement everyone already knows about. It is also the one that matters least on its own. Pitch alone is a poor predictor of how a voice is heard.
The shape of the sound
The big one. Resonance is your vocal tract colouring the tone — formants, in acoustic terms. It does most of the perceptual work, and it is far harder to measure cleanly than pitch.
How the folds vibrate
Light and quick, or heavy and thick. You hear it as the difference between a voice that sounds slight and one that sounds substantial — and it moves independently of pitch.
Movement over time
Not a single value — a shape. How much the pitch travels while you talk: bouncy or monotone. It only exists across a whole phrase, so it is measured across one, not sampled in a moment.
Plain loudness
The simplest dimension, and on the list for an unglamorous reason: if you do not account for it, loudness quietly skews every other reading. Boring, and load-bearing.
Prosody is a shape, not a number. Pitch traced across a whole phrase — the kind of contour the listening tools teach your ear to read.
Now the honest part.
A spectrogram will hand you a pitch to three decimal places. That precision is real, and it is also a trap — a number like that implies a voice can be 81% correct. It cannot. So Mira reports a range, not a score: approaching, in range, exploring.
“Exploring” is deliberate. It is not a polite word for wrong. The math can validate one thing — that a reading sits inside the range you are aiming at. It cannot validate a voice. So the feedback says only what the measurement can honestly support, and not one word more.
All of it runs on-device. The processing pipeline lives in your browser, which is what keeps the audio private — and also means the measurement is yours to inspect, not a verdict handed down from a server you cannot see.
No red. No number. No verdict.
Mira's interface follows three rules. No red — failure-coloured UI has no place in something you do every day, for yourself. No number — for the reason above. No verdict — the copy never tells you that you got it right; it tells you what it heard.
What is left is calm. Soft tones, generous space, and language that treats you as the expert on your own voice — because you are. It is built mobile-first, works offline once loaded, and is tested against WCAG AA.
A voice tool that never hears your voice — and never grades it.
Mira's analysis runs entirely on your device. Recordings are end-to-end encrypted with a key derived from your password — we could not listen to them if we wanted to. And nothing in the product assigns your voice a score.
It is open source under AGPL-3.0 and friendly to self-hosting, so the tool can outlast the people who started it. That is the whole point of building it this way.
Mira is in active development.
The Phase 1 release — practice, library, listening — is taking shape now. Have a product that needs to be private, patient, and genuinely yours?