Timbreloom

A local voice studio for your Mac.

Clone a voice. Keep a performance. Turn speech into text. Bring dialogue to picture.

Get started · User guide · Mac app · Architecture

Timbreloom brings four audio workflows into a native Mac app, with a shared local library and an English interface. Audio processing runs on your Mac using Qwen3-TTS, Whisper, Nemotron, Moonshine, and Seed-VC. Download the models you need once, then work offline.

Four ways to work with a voice

Workspace	What you can do
Voice cloning	Save a short reference recording and generate a sentence, a chapter, or a full script in that voice.
Voice conversion	Give an existing audio or video performance a new voice while preserving its rhythm, pauses, and expression.
Transcription	Turn audio or video into editable, timestamped text; export TXT, SRT, or VTT.
Dubbing studio	Assign roles and voices, edit lines on a shared timeline, regenerate individual segments, and export WAV, SRT, and MP4.

From a reference to a new script	From a recording to editable words

The screenshots show the actual Mac app using a separate demonstration library.

• Listen while it works. Finished speech segments join a continuous playback queue as generation progresses.
• Keep your work. Voices, transcripts, projects, and job progress are saved locally; supported jobs resume after a restart.
• Edit without starting over. Dubbing caches unchanged segments and lets you retry individual lines.
• Preserve every line. Safe duration fitting adjusts playback within 0.8x to 1.25x and reports remaining overlaps without cutting dialogue.
• Take projects with you. A .timbreloom bundle carries the project, referenced voices, media, and generated audio.

Get started

Timbreloom is built for Apple Silicon Macs. The current standalone app requires macOS 26.2 or later. This repository provides the source and build tools; the current local binary candidate is development-signed and has not been notarized for public distribution. See release validation for the exact status and testing limits.

Run from source

Install uv and FFmpeg, then clone this repository. The main environment uses Python 3.12; Seed-VC uses an isolated Python 3.10 environment, and the streaming transcription models use a separate Python 3.12 environment.

brew install uv ffmpeg
git clone https://github.com/Audiofool934/Timbreloom.git
cd Timbreloom
./setup.sh
./start.sh

Open localhost:7865. The developer setup downloads Qwen, Whisper, and Seed-VC and prepares the environments, so allow several gigabytes of disk space. Download the optional Nemotron and Moonshine models from Models. Models and your library are excluded from Git.

Build the Mac app

After setting up the project:

.venv/bin/python scripts/build_macos_app.py --install
open ~/Applications/Timbreloom.app

This builds a development app connected to your checkout. For a standalone app with bundled runtimes, see the Mac build guide. For an installed-app update, follow the update workflow, including version increments and removal of duplicate app entries.

Models, on your terms

The standalone app includes the runtime and downloads no model weights at launch. Open Models or press ⌘, to download, pause, or resume each model separately. Existing library items remain available to view and edit while a model is missing.

Model	Enables	Download size
Qwen3-TTS 0.6B Base, BF16	Voice cloning and dubbing	About 2.52 GB
Whisper Large v3 Turbo, Q4	Multilingual transcription with word timestamps	About 464 MB
Nemotron 3.5 ASR 0.6B, 8-bit	Multilingual streaming transcription	About 756 MB
Moonshine Small Streaming English	English streaming transcription	About 142 MB
Moonshine Tiny Streaming Mandarin	Mandarin streaming transcription	About 32 MB
Seed-VC and supporting models	Voice conversion	About 1.88 GB

Select a Transcription model for both imported files and live microphone input. Timbreloom remembers your choice, with Whisper as the initial default. Nemotron and Moonshine provide segment timestamps; Whisper provides word timestamps when available.

See the model manager and storage locations

The Mac app keeps models in ~/Library/Application Support/Timbreloom/models and your library in ~/Library/Application Support/Timbreloom/data. Replacing the app preserves both. Source runs use the repository's models/ and data/ folders by default. Set TIMBRELOOM_MODELS_DIR or TIMBRELOOM_DATA_DIR to use a different location.

Downloads use pinned revisions and verify every file with SHA-256 before enabling a model. Partial downloads are retained across pauses, connection failures, and app restarts.

Local by default

The service listens only on 127.0.0.1. Local voice cloning, conversion, transcription, and dubbing do not upload your recordings or generated audio. Microphone access is requested only when you start recording. The English interface supports multilingual speech and text.

Gemini is an optional text tool for punctuation, cleanup, and translation. It requires GEMINI_API_KEY and explicit confirmation before sending selected text; audio stays local. Cloud usage may incur charges. Keep credentials out of source files and Git.

Use your own voice or voices you have explicit permission to use, and disclose AI processing when sharing results.

Documentation and development

• User guide: the four workflows, long audio, exports, and recovery.
• Mac app: shortcuts, installation, builds, and validation.
• App update workflow: version increments, safe replacement, and a single app-search entry.
• Provider architecture: model interfaces, job persistence, and accelerator lifecycle.
• Project format: schema, segment caching, roles, and portable bundles.
• Mac design: visual system and interaction checks.
• Release validation: tested builds, evidence, and remaining limits.

TIMBRELOOM_DATA_DIR="$(mktemp -d -t timbreloom-tests)" TIMBRELOOM_PRELOAD=0 uv run pytest -q
node --test tests/test_browser_helpers.mjs
.venv/bin/python scripts/e2e_macos.py
.venv/bin/python scripts/e2e_macos.py --models

The WebKit checks use an isolated sample library and synthetic ASR. Real model validation is a separate step described in the Mac build guide.

License

The independently authored Timbreloom application and integration code is licensed under MIT. Third-party software and model weights retain their own licenses, including GPL-3.0 for the optional Seed-VC engine. See third-party notices and Seed-VC integration for details. Model weights, personal audio, local libraries, credentials, Python environments, and built applications are not included in this repository.