Privacy by architecture.Not merely by policy.
Cloud dictation services ask you to trust legal privacy policies while streaming your raw voice to remote servers. Murmur protects your confidential thoughts with physical architecture: your voice never leaves your device.
How voice flows from microphone to active window
Microphone → Your Computer (whisper.cpp & RAM) → Any App (Zero network egress)
1. Microphone
OS audio stream captured into temporary volatile RAM buffer (CPAL / WASAPI / CoreAudio).
2. Local Whisper Engine
Decoded on-device via whisper.cpp using native Metal (macOS) or DirectML/CUDA (Windows).
3. Context Rules
Local regex & jargon bias rules format text, strip fillers, and match app context.
4. Direct Insertion
Formatted text is typed directly into your focused cursor via native OS input injection.
The Murmur Local Data Boundary
A line-by-line breakdown of every data asset and its storage guarantee.
| Data Type | Storage Location | Cloud Transmission | Retention & Erasure | Audit Status |
|---|---|---|---|---|
| Microphone Audio | Volatile RAM buffer only | 0 Bytes (Never uploaded) | Instantly freed upon decode completion | Strictly Local |
| Transcripts & Pasted Text | Local SQLite database or RAM-only (Incognito) | 0 Bytes (Never uploaded) | Configurable auto-purge (0-30 days) or 1-click wipe | Strictly Local |
| Custom Dictionary & Snippets | Local SQLite (`dictionary` table) | 0 Bytes (Never uploaded) | Editable & erasable on demand | Strictly Local |
| App & Window Context | Ephemeral memory (Target app title/bundle ID) | 0 Bytes (Never uploaded) | Discarded after formatting rule evaluation | Strictly Local |
| User Account & Authentication | None (Zero accounts or logins required) | 0 Bytes (No user auth service exists) | N/A — 100% anonymous & local | Zero Accounts |
| Telemetry & Usage Analytics | None (Zero tracking SDKs) | 0 Bytes (No analytics beacons or crash telemetry) | N/A — No identifiers generated | Zero Telemetry |
| Whisper AI Model Weights | Local Disk (`~/.murmur/models`) | 1-time download from HuggingFace/GitHub | Permanent offline storage | Air-Gap Ready |
| Software Update Checks | None | GitHub Releases API (Query only) | Can be disabled via 'Check for updates' toggle | User Controlled |
Direct answers. Zero legalese.
Is audio ever uploaded? — No, never.
No. Never. Spoken voice captured from your microphone is streamed strictly into a temporary volatile RAM buffer on your physical machine and passed directly to the local Whisper inference engine. The millisecond decoding finishes, the raw audio buffer is immediately freed and zeroed. Zero bytes of audio are ever transmitted over the network.
Are transcripts ever uploaded? — No, never.
No. Never. Transcripts are generated locally on-device and typed directly into your active window at your cursor position via native OS keyboard injection. Murmur maintains no cloud transcript databases, remote backups, or synchronization servers. Your spoken words never touch an external server.
Does the product work with no internet connection?
Yes, 100%. Once model weights are stored on your local disk, Murmur requires zero internet connectivity to operate. You can dictate on airplanes, in air-gapped secure development facilities, or during network outages with zero drop in speed or accuracy.
What data leaves the device for licensing, updates, or error reporting?
Murmur is free and open-source under the MIT license, so there are zero licensing calls, seat tracking, or activation checks. For software updates, Murmur makes an optional, read-only HTTPS query to the official public GitHub Releases API to check version tags (which can be disabled with one click in Settings). Zero crash logs, stack traces, or error diagnostics leave your machine.
Is analytics disabled by default?
Yes. Telemetry and analytics are zero by default—Murmur contains no analytics SDKs, telemetry beacons, Google Analytics, PostHog, or third-party tracking scripts. We do not track words spoken, session frequencies, or what applications you dictate into.
Can users delete all local history?
Yes. Local history is stored in an on-device SQLite database (`murmur.db`). You can purge your entire history at any time with a single click in Settings, configure automatic purge retention rules (e.g. wipe after 24 hours, 7 days, or 30 days), or use Incognito Mode so transcripts are never written to disk in the first place.
What model runs locally?
Murmur runs open-source OpenAI Whisper architecture models compiled into compact, quantized GGML weights (Tiny, Base, Small, Medium, and Large-v3-Turbo) using whisper.cpp. Models execute natively with GPU hardware acceleration via Apple Metal on macOS and DirectML or CUDA on Windows.
Can advanced users block the app's network access and retain core functionality?
Yes. Advanced users and IT administrators can block Murmur in Windows Defender Firewall, Little Snitch, LuLu, or pf, or toggle Murmur's built-in 'Air-Gap Mode' in Settings. Core dictation, custom phonetic dictionary biasing, and context-aware formatting continue to function with 100% reliability.
How does Murmur differ from cloud dictation tools like Wispr Flow?
Cloud tools like Wispr Flow stream your voice over WebSockets to remote GPU clusters, store transcripts in cloud databases, and use cloud LLMs for post-processing. Murmur executes speech recognition and prompt formatting on your local hardware using whisper.cpp with sub-180ms latency.
How can I independently test and verify these claims?
You can monitor Murmur using standard packet capture tools (Wireshark, Little Snitch, LuLu, or Windows pktmon). When you press the dictation hotkey and speak for 10 minutes, your network monitor will register exactly 0 bytes of egress traffic.
Turn “Trust Me” into “Verify Me”
Don't take our word for it. Here is how security teams, enterprise compliance officers, and power users can independently verify Murmur with network inspection tools:
Real-Time Connection Monitoring
Launch LuLu or Little Snitch in Alert Mode. Start dictating in any application. Notice that zero connection alert prompts appear during voice capture, decode, or text insertion.
lulu --monitor /Applications/Murmur.appPacket Inspection During Dictation
Bind packet monitor to Murmur.exe. Dictate high-entropy paragraphs. Stop the trace and inspect the capture: zero TCP/UDP packets leave your network interface.
pktmon filter add -n murmur && pktmon start --etwProcess Bandwidth Auditing
Monitor process bandwidth in real-time. Dictate for several minutes continuously. Murmur will register 0 KB/s upload and 0 KB/s download.
sudo nethogs -d 1The Trust Wedge: Local Architecture vs. Cloud Promises
Cloud Dictation (e.g. Wispr Flow, Cloud APIs)
- Audio streams across public networks to remote GPU endpoints.
- Requires account creation, authentication tokens, and user IDs.
- Transcripts and voice snippets may be retained for model fine-tuning unless explicitly opted out.
- Dependent on third-party uptime, internet bandwidth, and corporate cloud retention policies.
Murmur On-Device Architecture
- 0 bytes of audio or transcript leave your physical hardware.
- Zero accounts, zero logins, zero marketing or crash telemetry.
- Model weights execute directly in local memory via whisper.cpp.
- Works completely offline on airplanes, air-gapped workstations, and strict enterprise networks.
Dictate with total confidence.
Murmur is free, open source, and built for people who value privacy as an absolute guarantee.