Problem Statement
People forget tasks, promises, meeting details, and daily instructions.
For elderly users or people with memory difficulties, missing those details can be much more serious.
Private edge memory assistant
Wearable Memory
for Real Conversations.
A private wearable assistant that saves conversations locally and answers questions about them later.
Capture
FBT-AS18
STT
Distil-Whisper
Memory
ChromaDB
Recall
Gemma-3
ACTIVE_LISTENING
100%
Local-first recall
Pi 5
Edge target
Voice
Owner gated
GPS
Context tags
Scroll the memory loop
01 / Capture
Speech becomes memory.
The headset captures useful speech and skips dead air.
neuraid.live/session
Current state
ACTIVE_LISTENING
VAD armed
Bluetooth trigger
Audio segment saved
Memory metadata
Time
10:42 AM
Place
CIS Dept
Owner query
What did we decide?Assistant
Store locally, verify the owner, retrieve memories, and speak back.Project Context
A private memory layer for everyday conversations.
NEURAID turns real conversations into private, searchable memory.
Proposed Solution
NEURAID captures useful speech, stores it locally, and lets the owner ask voice questions later.
It verifies the user, retrieves relevant memories, and speaks back a concise answer.
Use Cases
Built for hands-free recall.
Fast recall for conversations, classes, meetings, and care routines.
Recall conversations, tasks, and instructions.
Help students and professionals remember spoken details.
Support elderly users with memory difficulties.
Ask natural voice questions about stored memories.
System Workflow
From captured speech to spoken answer.
Six steps from speech capture to spoken answer.
STEP 01
Capture speech from a headset or microphone.
STEP 02
Transcribe useful audio locally.
STEP 03
Store searchable memory chunks in ChromaDB.
STEP 04
Verify the owner before recall.
STEP 05
Retrieve relevant memories for a question.
STEP 06
Generate and speak the answer locally.
Hardware Setup
Edge prototype stack
- Raspberry Pi 5 prototype.
- Bluetooth headset for hands-free control.
- Local storage for recordings and memory.
- Future version can move to compact wearable hardware.
Software Architecture
Local-first AI pipeline
- Local Distil-Whisper speech-to-text.
- Resemblyzer owner voice verification.
- mxbai GGUF embeddings with ChromaDB.
- Gemma 3 local answer generation.
- Piper TTS for spoken output.
Privacy & Security
Designed around controlled access to personal memory.
Local memory plus owner voice verification before recall.
Privacy Features
- Local memory storage.
- Voice authentication before access.
- Local answer generation.
- User-controlled reset and retention.
Current Prototype Limitation
Current prototype supports owner-versus-other attribution. Full multi-person diarization is a future upgrade.
Performance / Results
Working prototype validation
The Raspberry Pi 5 prototype demonstrates speech capture, memory storage, secure recall, and spoken answers.
- Speech detection, storage, retrieval, and TTS are working.
- Voice authentication gates private memory access.
- Gemma 3 answers from retrieved memory context.
- Prototype runs on edge hardware.
Measured values such as response time, transcription accuracy, authentication accuracy, and storage usage can be added after final testing.
Challenges Faced
Engineering constraints
- Keeping local AI fast on Raspberry Pi 5.
- Handling noisy real-world audio.
- Avoiding unnecessary recording.
- Keeping answers grounded in stored memory.
- Balancing privacy, consent, and retention.
Future Scope
Clear upgrades beyond the prototype.
The prototype proves the pipeline; the product can become smaller, smarter, and easier to wear.
Smaller wearable hardware.
Mobile app and memory timeline.
Improved speaker identification.
Multilingual support.
Caregiver and elderly-care features.
Real GPS integration.
FAQ
Evaluator questions, answered directly.
Is it cloud-based?
No. The core prototype uses local STT, storage, retrieval, LLM response generation, and TTS.
Can others access memories?
Voice authentication blocks recall when the speaker does not match the enrolled owner.
Can it work offline?
Yes, when the required local models are installed on the device.
Is this final hardware?
No. Raspberry Pi 5 proves the prototype; the product can later move to compact wearable hardware.
A Second Brain, Completely Offline.
A private edge-AI pipeline for capturing, searching, and recalling spoken memory.
Continuous Audio Capture
Headset audio capture with VAD filtering.
Local Speech-to-Text
Distil-Whisper turns speech into local text.
Owner-Gated Recall
Voice verification protects private recall.
Vector Retrieval (ChromaDB)
Stored memories become searchable context.
Gemma-3 Edge Inference
Local model answers from retrieved memories.
Instant Local TTS
Piper speaks answers back through audio.
Location Context
Memories can include place and time metadata.
SYSTEM_ARCHITECTURE
Multi-Threaded
Edge Pipeline.
Audio is processed locally, embedded with mxbai GGUF, and stored in ChromaDB.
On query, owner voice auth gates retrieval before Gemma-3 and Piper respond.
STAGE 01
Listen
Headset audio enters VAD and recorder workers.
STAGE 02
Transcribe
Distil-Whisper turns accepted segments into text.
STAGE 03
Remember
mxbai GGUF embeddings are stored in ChromaDB.
STAGE 04
Recall
Owner voice auth gates retrieval and Gemma-3 answers.
system@neuraid: ~/core
End-to-End Data Flow
Capture, memory, retrieval, response.
FBT-AS18 Headset
Bluetooth Audio Capture
Geo-Metadata
Resemblyzer
Voice Auth
Distil-Whisper
Local STT
ChromaDB
Vector RAG
Gemma-3
Local LLM
Piper TTS
Instant Audio Response
The Team Behind NEURAID
Final Year Project Group
Powered by the Edge
PythonCore Backend
Gemma-3Local LLM
ChromaDBVector Storage
Distil-WhisperLocal STT
mxbai GGUFLocal Embeddings
PiperLocal TTS
ResemblyzerVoice Auth
Linux evdevHardware I/O
PythonCore Backend
Gemma-3Local LLM
ChromaDBVector Storage
Distil-WhisperLocal STT
mxbai GGUFLocal Embeddings
PiperLocal TTS
ResemblyzerVoice Auth
Linux evdevHardware I/O