MamboTTS Desktop v1.3.3

Hebrew vocalization, from a tester's report.

Nikud is read, not thrown away

Vocalized Hebrew was stripped and handed to Renikud to guess the vowels, which
is the opposite of what writing the marks is for. Someone typing קֶלְוִין is
spelling out Kelvin precisely so the vowels are not guessed, and heard
קְלוֹוִין back.

RenikudPlus now separates the marks from the letters, feeds the plain letters to
the model so the model still supplies the stress, and decodes with the writer's
own vowels and consonant marks: dagesh, the shin and sin dots, holam and shuruk
vav. Words without nikud take the Renikud path exactly as before, so a document
mixing the two reads correctly throughout.

Measured through the phonemizer, with the unvocalized form as a control:

written	read as	unvocalized control
אֶן	ʔˈen	ʔˈan
קֶלְוִין	kelvˈin	kalvˈin
שָׁלוֹם	ʃalˈom	ʃalˈom

Also from the same report

A single letter followed by a geresh is the letter's name, so תשובה ב׳ is read
בֵּת rather than as the prefix be-.

A colon ends a clause, so the voice pauses there, except between digits where it
is a clock.

The editor takes Ctrl+Shift+R and Ctrl+Shift+L to force paragraph direction, the
way Word and Outlook do, for text the automatic guess gets wrong.

Downloads

platform	file
macOS, Apple Silicon	MamboTTS_1.3.3_aarch64.dmg
Windows, x64	MamboTTS_1.3.3_x64-setup.exe
Linux, x86_64	MamboTTS_1.3.3_amd64.AppImage, or the .deb or .rpm

Or install with one line: install.sh on macOS and Linux, install.ps1 on
Windows. All three platforms are built from the same commit.

MamboTTS Desktop v1.3.2

Hebrew numbers read the way Hebrew reads them

Windows only. GitHub Actions is disabled on this repository, so the release workflow could not build the macOS and Linux apps. macOS and Linux users should stay on v1.1.2 until those are attached.

Fixed

• Numbers in Hebrew (#8). 2011 was read as three numbers in a row, "shtayim, elef, ahat esre". A year, a counted noun, an ordinal in a date, a percentage, a sum and a range are different readings of the same digits, and Hebrew numbers agree in gender with the noun they count, so this is a lexicon problem rather than arithmetic. Numbers now go through heb-tts-normalizer, which maintains that lexicon: years, gender agreement, construct forms in the thousands, decimals, percentages, currency, ranges, and Hebrew calendar years. Two older bugs went with it: a comma was read as a decimal point (30,000 was "thirty point zero") and anything above 65,535 was left as bare digits.
• The speed slider offers 0.9x to 1.1x. The wider range from 1.3.1 measured as clean but did not hold up in use: further out, a take stops sounding like the same voice at a different pace. A speed saved under the old slider is clamped on load. The API keeps its 0.5 to 2.0 range for callers who want to experiment.

Install

irm https://github.com/maxmelichov/MamboTTS/releases/latest/download/install.ps1 | iex

Or run MamboTTS_1.3.2_x64-setup.exe directly. It installs for the current user, so nothing asks for an administrator, and upgrades an existing install in place. SmartScreen warns on first run because the build is unsigned. A fresh install downloads about 1.4 GB of models on first launch; after that everything runs offline.

MamboTTS_1.3.2_x64-setup.exe SHA-256: cceac8fc6a98116d1c8f0d80f99df466a2188d2599ee34f919edd1181e0e9b44

MamboTTS Desktop v1.3.1

Vocalized Hebrew no longer aborts, long text can no longer vanish, and the controls moved

Windows only. GitHub Actions is disabled on this repository, so the release workflow could not build the macOS and Linux apps. macOS and Linux users should stay on v1.1.2 until those are attached.

Fixed

• Vocalized Hebrew anywhere in the text used to abort the whole synthesis (#7). One vocalized poem or quotation threw away a forty-thousand-character document. No vocalized-text phonemizer ships by default, so the app now does what a reader who ignores the marks does: it strips the nikud and lets Renikud infer the vowels, the same as for the unvocalized text around it.
• Long text could come back far shorter than the input with no error (#9). Two causes, both closed:
  • A chunk whose phonemization came back empty was skipped silently. It now fails and names the text it could not speak, so a shortfall is an error rather than a "success" with most of the audio missing.
  • The Phonikud and phoneme-input paths sent the whole recording as a single frame, which the app refuses above 128 MB, about 25 minutes. Those paths did all the work and then failed at the end on anything long. They stream in 10-second pieces now, like the main path.

Changed

• Voice and speed moved next to the text, in the bar with the Generate button, since they are what you change between takes. Voice is a dropdown now. Language and the Hebrew speaker context stay in the sidebar, since they are set once.

Install

irm https://github.com/maxmelichov/MamboTTS/releases/latest/download/install.ps1 | iex

Or run MamboTTS_1.3.1_x64-setup.exe directly. It installs for the current user, so nothing asks for an administrator, and upgrades an existing install in place. SmartScreen warns on first run because the build is unsigned. A fresh install downloads about 1.4 GB of models on first launch; after that everything runs offline.

MamboTTS_1.3.1_x64-setup.exe SHA-256: aa4f9eb702c4e2ff4edc39ce0c76bb6d592dbedad502da663425c4a9f360cad8

MamboTTS Desktop v1.3.0

Speech speed, and an API button that actually stops the server

Windows only. GitHub Actions is disabled on this repository, so the release workflow could not build the macOS and Linux apps. macOS and Linux users should stay on v1.1.2 until those are attached.

New

• Speed control in the Studio, directly under the voice list. The engine re-paces the speech as it generates rather than stretching the finished audio, so pitch is unchanged. 1.0x reproduces the pace every previous version produced exactly. The range is 0.75x to 1.5x: slower than 0.75x the audio loses level and articulation rather than just slowing down, so the slider stops there. Your choice is remembered between launches.
• The app version is shown in the header, next to the model version. Before, the model version replaced it as soon as a model loaded, so there was no way to tell which build you were running.

Fixed

• The Start API button is a real toggle (#10). It used to call start in both states, so once it read "Running" a second press did nothing and the server could not be stopped from the app. Its label also came from a value captured when the panel opened, so it went on saying "Running" after the server had exited. It now reads the server's real state every few seconds.
• The model download size is stated correctly. Onboarding said ~570 MB; the real download is 1.4 GB.

Install

irm https://github.com/maxmelichov/MamboTTS/releases/latest/download/install.ps1 | iex

Or run MamboTTS_1.3.0_x64-setup.exe directly. It installs for the current user, so nothing asks for an administrator, and upgrades an existing install in place. SmartScreen warns on first run because the build is unsigned. A fresh install downloads about 1.4 GB of models on first launch; after that everything runs offline.

MamboTTS_1.3.0_x64-setup.exe SHA-256: 38176f7433de9ef3b1dc6d8a1c3f0200984af10e2745783424a4d9998727f739

MamboTTS Desktop v1.2.0

Superseded by v1.3.0. The installer attached here was replaced after this release was tagged with a later build that already contains the 1.3.0 speed control, so it does not exactly match the mambotts-desktop-v1.2.0 tag. Install v1.3.0 instead.

Windows desktop app, and five working languages

Windows gets a real desktop app for the first time. It was built, installed, and run on Windows 10 rather than reasoned about.

Windows only. This release does not yet carry macOS or Linux artifacts. GitHub Actions is disabled on this repository, so the release workflow could not build them. macOS and Linux users should stay on v1.1.2 until those are attached.

The bug that broke every non-Hebrew language

espeak-ng-data was never packaged into any artifact. Without it espeak-ng falls back to the data directory baked in when it was compiled, which is a path on the build machine, so it failed to initialise on every end-user machine and English, Spanish, German and Italian all returned HTTP 500. Only Hebrew worked, because it goes through Renikud instead. macOS and Linux had the same bug and get the same fix.

Also fixed

• Windows ships its own much older onnxruntime.dll in System32 for Windows ML, which shadowed the bundled 1.23 runtime and killed the sidecar inside the loader with 0xC000007B before it could write a word to stderr. The runtime now sits beside the executables, the only search location that wins.
• Tauri hands back \\?\ paths, and espeak-ng joins them with forward slashes, which Windows does not normalise inside such a path. The prefix is now stripped.
• --exit-with-parent is a no-op on Windows, so a crash or End Task left an invisible server holding the model in memory and locking its own executable against the next install. It is now confined to a job object with kill-on-close.
• Long text no longer fails. The finished recording was sent as a single frame against a 128 MB client ceiling, roughly 27,000 characters, and failed after the synthesis had already run to completion. It now streams in slices; verified at 35,000 characters producing 32.9 minutes of audio.

Install

irm https://github.com/maxmelichov/MamboTTS/releases/download/mambotts-desktop-v1.2.0/install.ps1 | iex

Or run MamboTTS_1.2.0_x64-setup.exe directly. It installs per-user, so nothing asks for an administrator. SmartScreen will warn on first run because the build is unsigned. On first launch the app downloads about 1.5 GB of models; everything after that is offline.

MamboTTS_1.2.0_x64-setup.exe SHA-256: 9c824f82652bc6e12b4da447cbfe300c5a3f69b055f50ce22fbd26144882857d

MamboTTS Desktop v1.1.2

BlueTTS only, and Linux joins macOS as a shipping platform.

Fixed

A fresh install could not generate audio. The studio asked for a voice named Rotem, which stopped existing when the catalog was renamed to Noa, Lily, Daniel, and Adam. Every first generation failed with "unknown Blue voice Rotem". The studio no longer keeps its own copy of the catalog; it adopts whatever the loaded model reports.

Changed

The Qwen3-TTS runtime and its GGML build path are gone. MamboTTS is BlueTTS only, which is what made the Linux and Windows builds practical.

Install

One command per platform.

macOS on Apple Silicon

curl -fsSL https://github.com/maxmelichov/MamboTTS/releases/latest/download/install.sh | sh

Linux on x86_64

curl -fsSL https://github.com/maxmelichov/MamboTTS/releases/latest/download/install.sh | sh

Windows on x86_64

irm https://github.com/maxmelichov/MamboTTS/releases/latest/download/install.ps1 | iex

Downloads

Platform	File
macOS, Apple Silicon	MamboTTS_1.1.2_aarch64.dmg
Linux, any distribution	MamboTTS_1.1.2_amd64.AppImage
Debian and Ubuntu	MamboTTS_1.1.2_amd64.deb
Fedora and RHEL	MamboTTS-1.1.2-1.x86_64.rpm
Windows, server and API only	MamboTTS-Server_1.1.2_x86_64-windows.zip

The macOS build is ad-hoc signed rather than Developer ID signed, so Gatekeeper will block it on first launch if you install the DMG by hand. Right-click the app and choose Open to get past that. install.sh does this for you: it clears com.apple.quarantine after copying the app, which is the same approval that dialog asks for.

There is still no Windows desktop app, because producing the installer needs WiX and NSIS on a Windows machine and this release was built on a Mac. What Windows now gets instead is the local server and its HTTP API, in MamboTTS-Server_1.1.2_x86_64-windows.zip. install.ps1 unpacks it into %LOCALAPPDATA%\MamboTTS, downloads the BlueTTS model and the Renikud phonemizer, and writes a launcher that starts the server and opens the Swagger page. That gives you /v1/audio/speech, /v1/voices, /v1/phonemize, and /skill, which is everything the desktop app uses to make audio.

install.ps1 has not been executed anywhere, since there is no Windows machine on the build host. Please open an issue if it fails for you.

MamboTTS Desktop v1.1.1

macOS (Apple Silicon) build.

Named voices. The BlueTTS 2.5 voices are now Noa, Lily, Daniel and Adam, matching the reference app. Noa is the default. The old file stems (libri_female_1088 and friends) still resolve, so a pinned voice keeps working.

Retired female.json. It is no longer downloaded or selectable, and the female1 alias is gone with it. All four remaining voices are LibriTTS English speakers: Hebrew text still synthesizes, but in an English speaker's timbre.

Everything from v1.1.0 also applies: the MamboTTS rename with bundle id com.maxmelichov.mambotts, BlueTTS 2.5 at 44.1 kHz under models/bluetts-2.5/, and macOS-only builds.

Unsigned. Not code-signed or notarized, so Gatekeeper will warn on first launch. Right-click the app and choose Open, or clear the quarantine flag:

xattr -dr com.apple.quarantine /Applications/MamboTTS.app

MamboTTS Desktop v1.1.0

macOS (Apple Silicon) build.

Renamed to MamboTTS. New bundle id com.maxmelichov.mambotts, so this installs alongside an older MamboRambo/MamboBlue copy rather than upgrading it, and starts with a fresh model directory.

BlueTTS 2.5. New checkpoint at 44.1 kHz, installed under models/bluetts-2.5/ so it does not collide with a v2 bundle. Speech in Hebrew, English, Spanish, German and Italian.

Voices changed. 2.5 ships one Hebrew speaker (Rotem, the same voice as before) plus four LibriTTS English speakers. There is no Hebrew male voice in this bundle, so Roi is not available.

macOS only. Windows and Linux are no longer built.

Unsigned. This build is not code-signed or notarized, so macOS Gatekeeper will warn on first launch. Right-click the app and choose Open, or clear the quarantine flag:

xattr -dr com.apple.quarantine /Applications/MamboTTS.app

MamboRambo Desktop v1.0.8

MamboRambo Desktop 1.0.8

Downloads

• macOS Apple Silicon: MamboRambo_1.0.8_aarch64.dmg
• Ubuntu/Debian x64: MamboRambo_1.0.8_amd64.deb
• Fedora/RHEL x64: MamboRambo-1.0.8-1.x86_64.rpm
• Windows x64 (portable): MamboRambo_1.0.8_x64-portable.zip

What's new

• Advanced IPA / phoneme editor mode
• Phonikud Hebrew G2P + interactive diacritics editing
• RenikudPlus model upgrade with speaker-context controls
• Docs for contributing open-source models

Notes

• Voices: Rotem & Roi
• Languages: Auto, Hebrew, English (plus model-supported extras)
• First launch downloads the Blue ONNX / RenikudPlus models from Hugging Face
• Windows portable build requires WebView2 Runtime

Built from ce17b02 on main.

MamboRambo Desktop 1.0.7

MamboRambo Desktop 1.0.7 (BlueTTS)

Downloads

• macOS Apple Silicon: MamboRambo_1.0.7_aarch64.dmg
• Ubuntu/Debian x64: MamboRambo_1.0.7_amd64.deb
• Fedora/RHEL x64: MamboRambo-1.0.7-1.x86_64.rpm
• Windows x64 setup: MamboRambo_1.0.7_x64-setup.exe
• Windows x64 MSI: MamboRambo_1.0.7_x64_en-US.msi

What's new

• Streaming playback: audio starts as soon as the first synthesis chunk is ready
• After generation finishes, the player uses the finalized WAV (no stale chunk noise)
• BlueTTS text/code handling parity improvements

Notes

• Voices: Rotem & Roi
• Languages: Auto, Hebrew, English
• First launch downloads the Blue ONNX model from Hugging Face

Built from 9d0e59b (streaming + Windows packaging fixes).