State-of-the-art on-device speech recognition
Speech recognition runs entirely on the Noter Vault device. We pair it with custom diarisation and speaker tooling that we have calibrated for confidential meeting capture.
Technical specifications
Private AI hardware for local transcription, notes, and Theia.
Noter Vault combines dedicated room hardware, on-device speech recognition with custom speaker tooling, custom-tuned Theia, and an AES-256-GCM encrypted device vault.
Spec summary
One appliance for the private meeting workflow.
Noter Vault is designed to capture the room, transcribe locally, structure the record, answer with Theia, and store the session in an encrypted local vault. No cloud account is required and no cloud upload is required for transcription, notes, or Theia.
Specifications
Technical summary
| Product category | Offline AI meeting recorder and private-AI appliance |
|---|---|
| Primary workflow | Capture, transcript, structured notes, Theia, encrypted vault, controlled export |
| ASR stack | State-of-the-art on-device speech recognition with custom diarisation and speaker tooling, calibrated for confidential meeting capture. |
| Theia | Theia is a high-capability local language model that we tune end-to-end for private meeting work — proprietary prompt engineering, dynamic prompting, temperature and context-window calibration, and response-shape testing, all running on a model that never leaves the device. |
| Vault encryption | AES-256-GCM encrypted device vault with configurable password protection. Vault keys are derived with PBKDF2-HMAC-SHA256 at 240,000 iterations. The vault encrypts every session artefact: metadata, transcript segments, structured note JSON, session snapshot, Theia conversation messages, and retained audio. |
| Storage claim | Up to 7,000 hours of raw audio or up to 1,000,000 completed notes with transcripts, whichever fills first. |
| Connectivity posture | No Wi-Fi hardware in production devices. |
| Cloud account | Not required |
| Cloud upload for transcription, notes, and Theia | Not required |
| Telemetry | None at all |
| Export formats | PDF, DOCX, TXT |
AI stack
Local transcription and local assistant
The public model description should stay useful without disclosing deployment-sensitive internals.
Custom speaker tooling
On-device diarisation and speaker-aware transcript evidence help turn a private room conversation into a structured record, with no audio leaving the device.
Custom-tuned Theia language model
Theia is a high-capability local language model that we tune end-to-end for private meeting work: proprietary prompt engineering, dynamic prompting, temperature and context-window calibration, and response-shape testing — all running on a model that never leaves the device.
Structured note generation
Templates, meeting context, glossary terms, and local transcript evidence help produce focused meeting records instead of generic notes.
Device and vault
Hardware and storage posture
Dedicated private-AI appliance
The product is room-owned hardware, not an external meeting attendee or browser-based cloud scribe.
No Wi-Fi hardware
Production devices are built without Wi-Fi chips, reinforcing the local-first security model.
Encrypted device vault
Records are stored in an AES-256-GCM encrypted device vault with configurable vault password protection.
Large local capacity
Publishable capacity is up to 7,000 hours of raw audio or up to 1,000,000 completed notes with transcripts, whichever fills first.
The system
Engineered for evidence, not just transcription
Noter Vault is more than offline speech-to-text. The appliance runs a coherent meeting engine: dynamic recognition, split-process capture, transcript hygiene gates, evidence-bound outputs, session-grounded Theia, cryptographic vault hardening, and a deterministic clean boot. Each layer is purpose-built for confidential meeting work, and together they are what separates Noter Vault from a generic transcription wrapper.
Evidence-bound, not improvised
Theia structured outputs are grounded to the local transcript. Templates source-quote the underlying record, meeting content is fenced so it cannot be misused as model instructions, and unsupported answers are flagged rather than fabricated. Outputs cite the room, or they say so.
Session-grounded Theia
Theia answers only from the meeting record loaded in the vault. It does not reach beyond the session for context. Where the session does not contain the evidence, Theia returns an explicit unsupported response rather than improvising.
Transcript hygiene safeguards
Multiple on-device gates strip prompt leaks, suppress duplicates, reject noise artefacts, and discard subtitle and caption junk before any transcript reaches the structured note. The record that enters the vault is the meeting, not the model chatter.
Recognition tuned to the meeting
Before capture starts, Noter Vault primes its on-device speech recognition with the meeting context: glossary terms, expected speakers, hotwords, and prior session continuity. Recognition adapts to the room, not a generic vocabulary.
Split capture and inference
Audio capture and AI inference run as separate on-device processes. The capture pipeline buffers and chunks audio independently of the recognition workload, so the device stays responsive throughout long meetings.
Clean execution state per session
Every session starts from a deterministic local state. The device syncs storage, drops page cache, and cycles swap before a recording begins, so each meeting runs in a clean execution environment with no residue from prior sessions.
Local speaker profiles
Teams can enrol speaker profiles on the device. Combined with on-device diarisation, a structured transcript labels who said what, with voiceprints that never leave the appliance.
Cryptographic vault hardening
The encrypted device vault uses AES-256-GCM with PBKDF2-HMAC-SHA256 key derivation at 240,000 iterations. Every session artefact is encrypted: metadata, transcript segments, structured note JSON, session snapshot, Theia conversation, and any retained audio.
Software workflow
Capture, records, Theia, and export in one desktop product.
The desktop workflow includes setup, meetings, vault, Theia, people, glossary, settings, transcript evidence, session snapshots, saved-session Theia, and PDF, DOCX, or TXT exports.
FAQ
Questions buyers ask
What speech recognition does Noter Vault use?
Noter Vault uses state-of-the-art on-device speech recognition with custom diarisation and speaker tooling, calibrated for confidential meeting capture. Audio never leaves the device.
What is Theia?
Theia is the local private assistant inside Noter Vault. It can answer questions, summarise, draft, and analyse the selected meeting record without uploading the transcript to the cloud.
How much can Noter Vault store?
The production claim is up to 7,000 hours of raw audio or up to 1,000,000 completed notes with transcripts, whichever fills first. Actual capacity depends on configuration and retained audio policy.
Can Noter Vault be used without an account?
Yes. Noter Vault does not require a cloud account for setup or use, and it has no telemetry at all. The device shares no information with anyone, not even Noter, unless the user explicitly chooses to export or share it.
Book a technical fit demo.
Noter Vault is configured around your industry, meeting types, templates, ASR context, glossary, and deployment needs.