Measure what your voice is actually doing, and watch it change over months.
Pitch, resonance, weight and voice quality — live while you speak, and tracked across sessions.
Live view | Guided run | Session list |
Column menu | Recording in detail | Settings · Analysis |
Settings · Target profiles |
Live view — everything in one bar: language, microphone, recording type, target voice, guided run. Below it the readouts, the zone bar, the spectrogram with F1 and F2 marked, the pitch history and the practice text.
Guided run — a panel inside the Live tab rather than a window, so the spectrogram keeps running while you hold a sound.
Session list — right-click a row for details, playback, renaming, deletion or changing the recording type afterwards. Recordings without a usable voice are greyed out and say so instead of showing invented numbers. Right-clicking a header offers sorting, moving, hiding and all 23 columns as a checkable list.
Detail view — eighteen metrics, each with an explanation behind the ⓘ, checked against a target profile and optionally compared against another recording. "Advanced" unfolds the waveform with a draggable region. The example shows a recording that was too quiet: the metrics stay blank rather than being guessed.
Settings — analysis parameters with templates on one tab, target profiles on the other, where each of the eighteen metrics can be switched on and given a range.
This is a measuring instrument, not a therapy programme. It tells you what your voice is doing. It does not tell you what to do about it, and it cannot hear whether you sound good.
- Never train through pain. If your throat hurts, if your voice feels scratchy, or if you are hoarse after a session, stop. Those are signs of too much tension or pressure, not signs of effort paying off. Damage from pushing a voice takes far longer to undo than it takes to cause.
- The numbers are not a score. The reference ranges are population averages from the phonetics literature. Two people with identical measurements can be perceived completely differently, and perception depends on far more than anything measurable here. Use them for orientation and to compare yourself against your own earlier recordings.
- This cannot diagnose anything. Jitter, shimmer and harmonicity appear in clinical literature, but the values here come from ordinary speech on consumer microphones and mean nothing diagnostically. Persistent hoarseness belongs to an ENT doctor.
- A speech-language pathologist beats any software. A few sessions with someone who can actually hear you will save you hundreds of hours of guessing, because early on you cannot judge your own output reliably. In Germany, gender-affirming voice therapy is prescribable and covered by statutory health insurance.
Community resources worth knowing: the Voice Resource Project wiki, Spectrus for a quick browser spectrogram, and r/transvoice.
Live view. Pitch (F0), the first two formants, weight (H1–H2) and the input level as large readouts. Pick a target voice and they turn green or amber against it; leave it off and they stay neutral. A scrolling spectrogram with dashed markers where F1 and F2 sit, so you can see resonance move while you speak. A pitch history over the last 30 seconds with your target zone shaded. A zone bar showing where you are between your configured boundaries.
Sessions. Record, and every take is stored as a WAV plus a set of metrics. The list shows date, duration, level, pitch and formants at a glance. Anything that was too quiet or contained no usable voice is marked as such instead of being given misleading numbers.
Detail view. Eighteen metrics per recording, each with an explanation behind an ⓘ. Pick a target profile — masculine, androgynous or feminine — and every metric is checked against the corresponding range. Pick any other recording as a comparison and get a delta column. Export the whole thing as readable text or as CSV.
Advanced mode. The waveform of the recording with a draggable region. Select part of it, analyse just that, play just that. This is how you get meaningful formant, jitter and shimmer numbers: record a long take with a sustained vowel, then cut out the stable middle.
Recording types. Every take is filed as a reading text, a pitch test (a held hum), /a/, /i/, /u/ or free. Picked before recording and sortable as a column, so a held vowel is never averaged together with connected speech.
Guided run. A toggle, off by default, that shows a slim strip in the Live tab — not a window in your face. It walks you through hum, /a/, /i/ and /u/ with a countdown, trims each take to its steadiest stretch and files it, while the readouts and spectrogram keep running beside it. This is the step that makes formants, jitter and shimmer comparable between sessions, and the one people otherwise skip.
Session list. Click a column header to sort, click again to reverse. Right-click a row for details, playback, rename or delete; right-click a column header to sort, hide it or toggle any column straight from the menu. Columns can be dragged into any order. The view dialog picks which of the 23 available columns to show — including H1, H2, jitter, shimmer and voice breaks — sets the sort order and optionally limits the list to a date range. The list exports to text or CSV.
Debug window. Reachable from the settings dialog. Warnings and caught exceptions with tracebacks and environment details, exportable as a text file for bug reports. Empty is the normal state.
Target profiles. A dropdown in the Live toolbar picks the target voice — masculine, androgynous, feminine or one of your own. It starts at no target on purpose: a colour changing while you speak pulls your attention to the screen, and attention on the screen is attention off your voice. The detail view has its own selection, defaulting to feminine, so you can look at targets afterwards without them nagging you during practice. The settings have a tab for editing the range of any of the eighteen metrics — H1 and H2 separately included — and saving it under your own name; the built-in three start from population averages and can be adjusted, saved under a new name, or reset back to the literature values.
Themes. Eight built-in colour schemes, every colour role recolourable individually, and an optional background image with adjustable panel transparency.
Settings. Every analysis parameter is adjustable while the program runs, with six built-in templates for different situations. Nothing requires a restart.
Detachable tabs. Right-click the tab bar or double-click a tab to pull it into its own window — useful for watching the live readouts on one screen while browsing sessions on another. Closing the window docks it back.
| Metric | What it tells you |
|---|---|
| F0 median | Average speaking pitch. The most cited value and the least important one on its own. |
| F0 10 % / 90 % | How far down sentence endings drop and how far up you reach. |
| Intonation width | Standard deviation in semitones — how much melody your speech carries. |
| Pitch range | Distance between the two ends, in semitones. |
| F1 / F2 / F3 | Formants. F2 is the single most useful number for resonance; F3 tracks vocal tract length closely. |
| H1 / H2 / H1–H2 | Levels of the first two harmonics and their gap. A large gap means light, breathy production; a small one means heavy, pressed. |
| Clarity (HNR) | Harmonics-to-noise ratio. Low means breathy — or that the recording was too quiet. |
| Jitter / shimmer / voice breaks | Cycle-to-cycle stability. Reference ranges apply to sustained vowels only; connected speech is naturally far higher. |
| Voiced share | How much of the recording was recognised as voice at all. |
| Recording level | Peak level. Below −40 dBFS the measurements stop being trustworthy. |
Resonance matters more than pitch. Raising F0 without moving the formants produces a voice that reads as low-pitched rather than differently gendered, which is why F2, F3 and H1–H2 are given as much weight here as F0.
curl -fsSL https://raw.githubusercontent.com/yakuda-stack/Dream-VoiceTraining/main/install.sh | bashDetects Debian, Ubuntu, Fedora, Arch, CachyOS and openSUSE, installs the
system packages that have no Python equivalent (PortAudio, venv, git), fetches
the source into ~/.local/lib/dream-voicetraining, builds its own Python
environment and adds a menu entry. sudo is used only for the system packages
and is announced before it happens; everything else stays under ~/.local.
Piping a script into a shell means trusting it. If you would rather look first, which is the sensible habit:
curl -fsSLO https://raw.githubusercontent.com/yakuda-stack/Dream-VoiceTraining/main/install.sh
less install.sh
bash install.shRemoving it again keeps your recordings and settings:
bash install.sh --uninstallUseful flags: --no-deps skips the system packages, PREFIX=/somewhere moves
the installation.
praat-parselmouth is not in the AUR yet, so it is built first. This compiles
the bundled Praat sources and takes a few minutes.
git clone https://github.com/yakuda-stack/Dream-VoiceTraining
cd Dream-VoiceTraining/packaging
makepkg -si -p PKGBUILD.python-praat-parselmouth
paru -Uiparu rather than makepkg for the second step: python-sounddevice lives
in the AUR, and makepkg only knows pacman. With yay: yay -Bi .
If building Praat fails after a Python version bump in Arch, use
PKGBUILD.python-praat-parselmouth-bin instead — it installs the official
wheel and needs no compiler.
Download from Releases:
chmod +x Dream-VoiceTraining-*.AppImage
./Dream-VoiceTraining-*.AppImageOne file for every system: the runtime uses FUSE 3, falls back to FUSE 2 and extracts itself if neither is available. Nothing to install.
PortAudio has to be present on the host — it is not bundled, because audio routing has to come from your system to work at all.
python -m venv .venv
source .venv/bin/activate # fish: source .venv/bin/activate.fish
pip install -r requirements-dev.txt
python main.pyPortAudio is required and is not a Python package:
sudo pacman -S portaudio # Arch, CachyOS
sudo apt install libportaudio2 libpulse0 # Debian, Ubuntu, Mint
sudo dnf install portaudio pulseaudio-utils # Fedora
sudo zypper install portaudio pulseaudio-utils # openSUSEpactl (from libpulse) is optional but strongly recommended. Without it,
device names fall back to raw ALSA strings like
USB Audio Device: USB Audio (hw:1,0) instead of the readable PipeWire
descriptions.
A virtual environment does not survive its parent folder being renamed or
moved — the activation script writes absolute paths. If you move the project,
delete .venv and create it again.
Fix the level before anything else. Open pavucontrol, go to the
recording tab and raise the input gain until the level card sits around −20 dB
while you speak normally. Sit roughly a hand's width from the microphone,
slightly off to the side so plosives do not hit it directly. If the level card
turns red, the signal is below the silence threshold and nothing is being
analysed.
Read the practice text once, all of it, and record it. That is your baseline.
Note the numbers. Then record /a/, /i/ and /u/ held for three seconds
each, and use advanced mode to cut out the stable middle of each — those give
you formant values that will still be comparable in two months.
From then on, ten to fifteen minutes a day beats two hours on a weekend. Use the same text, the same microphone position and the same template every time, otherwise you are measuring your setup rather than your voice.
Do not read anything into differences between two takes recorded minutes apart. Day-to-day variation, microphone distance and how warmed up you are move these numbers more than weeks of practice do. Trends need weeks.
Follows the XDG Base Directory Specification:
~/.config/dream-voicetraining/config.json settings, templates
~/.local/share/dream-voicetraining/sessions/ recordings and metrics
DREAM_VOICETRAINING_HOME overrides both with a single directory, which is
useful for testing and for portable setups. Data from earlier versions is
picked up automatically on first start; existing files are never overwritten.
Everything is plain WAV and JSON. Nothing is uploaded anywhere, there is no telemetry, and the program makes no network connections at all.
diag.py records three seconds through exactly the same path the program uses
and prints what arrives at the analysis stage:
python diag.py --list
python diag.py --device 0Check "share exactly 0", "distinct values" and the example spectrum column at the end. That output turned a two-hour guessing session into a five-minute fix twice during development.
Common cases: a red level card means the input gain is too low or you picked a monitor source instead of a microphone. "No speech signal" on a recording where you definitely spoke means the level was too low for the voicing threshold — fix the gain rather than lowering the threshold.
Every release is documented in CHANGELOG.md, including what was broken and why.
The step-by-step release workflow — GitHub, AppImage, AUR — is in update_DreamVoiceTraining.txt.
pip install -r requirements-dev.txt
python -m pytest tests/ -qOne hundred and two tests covering the analysis core, settings persistence, path resolution and migration, device enumeration and translation completeness. Two of them are regression tests for real bugs: noise being reported as a voice, and the spectrogram collapsing to one colour.
Adding a translated string means adding it to i18n.py and using t("key").
The test suite fails if a key is used but not defined, if a string is missing
either language, or if placeholders differ between the two.
Contributions welcome, especially from people who actually train their voice and can say which numbers turned out to be useful and which are noise.
- Discord — https://discord.gg/UkhJSz3Ctf
- Ko-fi — https://ko-fi.com/yakuda_ (entirely optional; the program is free software and stays that way)
Bug reports go to the issue tracker. The settings dialog has an About window with a button that copies your system information, and a debug window whose log can be exported — attaching both makes a report far easier to act on.
GPL-3.0-or-later. Not a free choice: this links Praat through Parselmouth, and Praat is GPL-3.0. See THIRD_PARTY_NOTICES.md for the full dependency list and the note on the LGPL components.
If you publish measurements made with this, cite Praat and Parselmouth rather than this tool.
Parts of this code were written with AI assistance. Review, testing and maintenance are mine. Every bug fixed in this repository was found by measuring, not by guessing, and the diagnostic script and the test suite exist so that stays true.