VoxBridge Compose is a local-first voice workspace for Windows and Linux. It captures speech, transcribes it on-device through the VoxBridge runtime, and can refine the result with an ordered set of local agents.
Important
This project is an independently maintained derivative of FOSS Voquill by Jack Brumley and its contributors. Voquill provided the application foundation and remains a major source of inspiration. VoxBridge Compose is not affiliated with or endorsed by the original maintainers. Please consider supporting the original project.
- Faster Whisper/CTranslate2 speech recognition by default, with whisper.cpp/Vulkan retained as the broad-hardware compatibility fallback
- Side-by-side raw transcript and refined text with independent scrolling
- Embedded local text refinement or an Ollama server on the local network
- Ordered, editable agent profiles with fidelity controls and optional context
- Private mode that retains no new history, recordings, or session logs
- Configurable offload locations for text and optional source audio
- Model warmup, progress reporting, diagnostics, hardware details, and statistics
- Windows taskbar activity indicators for recording and transcription
No account or hosted speech service is required.
Hover over the active agent to review its prompt and processing settings.
VoxBridge Compose keeps speech processing local and separates product workflow from inference-runtime details:
- Compose application: owns microphone capture, utterance ordering, continuous-recording policy, raw and refined documents, agents, history, offloading, status, and desktop behavior.
- VoxBridge runtime: presents a stable Rust-facing boundary over backend discovery, hardware capabilities, model lifecycle, warmup, cancellation, caching, fallback, and normalized results.
- Established inference backends: Faster Whisper/CTranslate2 is the default, efficient speech-recognition path using CUDA or optimized processor inference; whisper.cpp remains the compact, broad-hardware compatibility fallback. Embedded llama.cpp or a user-selected Ollama server performs text refinement.
Audio is not sent to Ollama. Ollama receives text only when it is selected as the refinement provider. Users of a network Ollama server are responsible for the privacy and security of that connection.
Live refinement uses a confirmed-prefix/revisable-tail pipeline: completed text remains stable while one recent tail is refined serially, and newer speech is coalesced for the following pass. Results are applied only to the exact source span that produced them, so a late response cannot overwrite newer dictation. Stopping recording waits for outstanding recognition and runs a separate full-document reconciliation. Output safeguards reject repetition, missing or reordered sentences, context leakage, malformed joins, tiny sentence fragments, paragraph collapse, and unsupported rewrites. This design is informed by the LocalAgreement research behind Whisper-Streaming and SimulStreaming; the implementation is original to VoxBridge Compose. See NOTICE.md for full credit and licensing context, including the earlier incremental-recognition stability work and bounded-revision research that also informed the design.
VoxBridge is deliberately not a universal AI-provider gateway. Projects such as LiteLLM and LocalAI already serve that category. Its narrower purpose is to keep the local speech-recognition and refinement backends required by Compose behind one coherent lifecycle and result contract. VoxBridge reuses inference projects; it does not claim to implement Whisper, CTranslate2, or llama.cpp inference.
VoxBridge Compose is under active development. Windows is the primary tested platform. Linux support is retained, but release packages should be treated as pre-release until exercised on representative systems.
VoxBridge Compose retains FOSS Voquill's original desktop foundation—including audio capture, global shortcuts, local history, diagnostics, model management, and cross-platform integrations—while introducing the following:
- A Compose-first workspace with equal, independently scrolling raw-transcript and refined-text panes
- Recording, readiness, active-agent progress, and profile controls in one compact header
- Follow-output scrolling that pauses when the user scrolls upward
- Revision history, reversion, copying, export, and configurable offloading
- A maximized horizontal layout and VoxBridge Compose visual identity
- Faster Whisper/CTranslate2 speech recognition by default, with whisper.cpp available through the VoxBridge runtime as the compatibility fallback
- Automatic CPU and graphics capability detection with appropriate engine selection, graphics acceleration, and CPU fallback
- Model download, initialization, warmup, transcription, and error progress
- Start and stop recording directly from Compose, gated until the complete processing pipeline is ready
- Press-to-toggle, hold-to-record, and continuous listening workflows
- Continuous-speech segmentation with configurable silence timing and retained context between utterances
- Microphone selection, sensitivity adjustment, and microphone testing
- Windows and Linux global-shortcut and permission integration
- Embedded in-process refinement through VoxBridge and llama.cpp
- Ollama refinement on the same computer or another machine on the local network; Ollama receives text, not recorded audio
- Ordered agents with names, prompts, fidelity safeguards, speed settings, enable/disable controls, and reusable profiles
- Comprehensive Rewrite as the default profile, including contextual correction of likely speech-recognition errors
- Sequential multi-agent execution, per-stage progress, acceptance safeguards, recomputation, and earlier-segment correction
- Serialized, generation-aware live refinement that safely coalesces new speech and discards stale results after provider changes
- Final whole-document reconciliation that preserves established paragraphs and stable text while cleaning the most recent dictation
- Optional bounded saved-history context per agent
- Qwen3 4B Instruct as the recommended embedded model, with a lighter Qwen2.5 1.5B option and support for a custom GGUF file
- Model installation state, download controls, preload progress, and a graphics-memory capacity check before managed downloads
- Safe switching between embedded refinement and local or network Ollama, including preload and automatic document reconciliation after the new provider becomes ready
- Context-aware filename suggestions based on the overall document rather than simply copying its opening words
- Private mode that prevents new history, diagnostic session logs, and recording WAV files from being saved
- Configurable history and recording-log retention, including forever for history; WAV recording logs remain disabled by default
- A default offload folder plus multiple saved destinations selectable from the Compose toolbar
- Session-specific destinations with optional remembered selection
--offload-location <path>and--offload-location=<path>launch arguments for a temporary custom destination without replacing the saved default- Native folder selection and offloads containing raw transcript, refined text, and optionally the matching source WAV
- Organized General, Text refinement, Agents, Speech recognition, Audio, and Data and logs settings
- Processor, graphics adapter, system memory, graphics memory, Vulkan, request, word, agent, timing, cache, and acceptance reporting
- Live graphics-memory use when the platform reports it, plus separate recognition and refinement estimates; local Ollama memory is read from its runtime when available and network Ollama is clearly marked as unmeasured
- Privacy-conscious bug-report preparation with local-path redaction and a review step before opening GitHub
- Windows taskbar indicators for recording and transcription activity
- Installed/latest version information, changelog access, and repository and release links, with optional automatic startup checks
- Destructive reset with an explicit warning covering downloaded models, custom agents, history, recordings, logs, saved locations, and local settings
- About workspace with hardware information, project/support links, license details, bug reporting, and prominent FOSS Voquill attribution
Set a destination for one session without changing the saved default:
"VoxBridge Compose.exe" --offload-location "D:\Transcripts\Session A"
voxbridge-compose --offload-location=/mnt/transcripts/session-a
Both --offload-location <path> and --offload-location=<path> are supported.
The launch argument takes priority for that session only.
- Whole-pipeline profiles
- Drag-and-drop agent ordering
- GitHub profile import and updates
- Per-agent model and provider selection
- Reviewable agent-to-agent prompt adaptation, allowing designated agents to propose improvements to later agents' prompts using chain results and feedback. Changes must be bounded, versioned, reversible, auditable, and require explicit user approval before becoming persistent defaults.
- User-facing rerun and cancellation controls
- Editable raw-transcript and refined-text panes
- Draft-and-apply saving for agent edits
- Complete portable packaging for the supported Faster Whisper/CTranslate2 backend, including managed runtime/model downloads, CUDA and processor capability detection, progress/cancellation, and safe utterance-boundary switching. Retain whisper.cpp/Vulkan as the compact, broad-compatibility default.
- Bundled Space Grotesk typography
- Removal of dormant legacy cloud and API code
- Comment, naming, encoding, and unused-code cleanup
- Broader Linux package testing
See CHANGELOG.md for release-by-release details.
Clone recursively so the separate VoxBridge runtime and its native engine submodules are available:
git clone --recurse-submodules https://github.com/tednv/VoxBridge-Compose.git
cd VoxBridge-Compose
npm ci
npm run tauri:build
See docs/BUILD.md for platform prerequisites and troubleshooting.
If VoxBridge Compose is useful to you, you can buy me some LLM tokens. GitHub also displays this link through .github/FUNDING.yml.
Support for the project that inspired this work belongs here too: donate to FOSS Voquill.
VoxBridge Compose gratefully acknowledges:
- FOSS Voquill, its creator Jack Brumley, and all contributors, for the original AGPL-licensed application, desktop integrations, and cross-platform foundation.
- VoxBridge, the separately maintained MIT-licensed local runtime used for speech recognition and embedded refinement.
- whisper.cpp and llama.cpp, which provide the native inference foundations used by VoxBridge.
- Qwen3 4B Instruct 2507, distributed as GGUF by Unsloth, the default embedded refinement model, and Ollama as an optional user-configured local model service.
- The Tauri, Preact, Tabler Icons, Rust, and broader open-source communities.
Original work retains its original copyright and license. Subsequent changes and new components retain the copyrights of their respective contributors. See NOTICE.md and THIRD_PARTY_NOTICES.md.
VoxBridge Compose is distributed under the GNU Affero General Public License version 3. The separate VoxBridge submodule is distributed under its own MIT license.