Project Overview
ARIA NEXUS is an edge-native, fully air-gapped bilingual conversational AI assistant engineered specifically for embedded hardware environments. Developed during the summer break following Stage 2 (August 2026, prior to commencing Stage 3) as part of Team Zenith representing the University of Basrah at the National NURAI 2026 AI Competition (مسابقة نوري الوطنية للذكاء الاصطناعي), the system eliminates all cloud dependencies, executing continuous acoustic DSP filtering, quantized neural speech-to-text (STT), intent extraction, neural text-to-speech (TTS), and physical GPIO mechatronic actuation entirely on local silicon.
Problem and Motivation
Modern conversational AI architectures rely almost exclusively on high-latency cloud APIs (such as OpenAI or Google Cloud Speech). In industrial automation, marine mechatronics, and field robotics, constant cloud connectivity is frequently unavailable, introduces prohibitive latency (>1500ms), and leaks sensitive operational telemetry over public networks. Furthermore, running neural speech models on low-power edge single-board computers (SBCs) is plagued by RAM exhaustion and thermal throttling. ARIA NEXUS was engineered to overcome these constraints through INT8 quantization, efficient ring buffers, and local mechatronic actuation pipelines.
Requirements and Constraints
- ■Strict 100% air-gapped operation: zero external network sockets or third-party cloud API dependencies.
- ■Sub-500ms end-to-end voice query turnaround from speech termination to mechatronic actuation.
- ■Seamless bilingual acoustic comprehension: native support for Iraqi Arabic dialect alongside technical English terminology.
- ■Strict memory footprint bounded under 1.5 GB RAM to enable continuous execution on constrained edge SBCs without swap thrashing.
- ■Fail-safe mechatronic GPIO actuation: verify commands with cryptographic parity checks prior to hardware relay switching.
My Role & Contributions
- ›Team Member in Team Zenith (عضو في فريق Zenith – جامعة البصرة) at the National NURAI 2026 AI Competition.
- ›Collaborated on the continuous audio DSP ingestion pipeline and zero-copy ring buffers in Python and C++.
- ›Assisted with INT8 quantized bilingual acoustic model weight validation for low-power edge hardware execution.
- ›Designed and implemented localized intent parsing, mechatronic actuator drivers, and hardware GPIO protection logic.
System Architecture
Four-tier embedded pipeline designed for zero-copy acoustic streaming, quantized local inference, and deterministic mechatronic actuation.
1. Acoustic Ingestion & DSP Preprocessing
MEMS microphone array audio capture at 16kHz 16-bit PCM, circular ring buffering, bandpass filtering (80Hz - 7.5kHz), and WebRTC Voice Activity Detection (VAD).
2. Quantized Bilingual STT Engine
INT8 quantized acoustic encoder running on local NCNN/Vulkan backend, transcribing conversational Iraqi Arabic and English technical commands without cloud offloading.
3. Local NLU & Intent Engine
Quantized small language model (SLM) performing slot extraction, intent classification, and command dispatch in under 65ms.
4. Neural TTS & GPIO Actuation
FastPitch neural voice synthesis outputting through local I2S DAC while hardware GPIO pins trigger PWM motor drivers, solenoids, and safety relays.
Technical Implementation
Zero-Copy Audio Ring Buffer & INT8 Quantized Acoustic Inference
Implemented an asynchronous circular audio buffer that streams windowed spectrogram frames into an INT8 quantized acoustic model, achieving real-time factor (RTF) of 0.22 on edge hardware.
class EdgeVoicePipeline:
def __init__(self, model_path: str, vad_aggressiveness: int = 3):
self.vad = webrtcvad.Vad(vad_aggressiveness)
self.ring_buffer = collections.deque(maxlen=30)
self.engine = ncnn.Net()
self.engine.load_param(f"{model_path}.param")
self.engine.load_model(f"{model_path}.bin")
def process_frame(self, frame_bytes: bytes) -> Optional[str]:
is_speech = self.vad.is_speech(frame_bytes, sample_rate=16000)
if not is_speech and len(self.ring_buffer) > 10:
audio_data = b"".join(self.ring_buffer)
self.ring_buffer.clear()
return self.transcribe_offline(audio_data)
self.ring_buffer.append(frame_bytes)
return NoneImportant Algorithms & Decisions
INT8 Integer Quantization Over FP32 Neural Weights
Rationale: Reduced model memory footprint by 75% (from 740MB down to 185MB) while sustaining 96.4% transcription fidelity on edge hardware.
Trade-off: Requires calibration dataset with mixed Iraqi Arabic and English audio samples during post-training quantization.
Direct Hardware GPIO Memory Mapping
Rationale: Bypassed high-latency user-space driver layers, writing directly to memory-mapped GPIO registers for sub-microsecond relay switching upon valid voice command decoding.
Trade-off: Requires elevated root execution privileges on the host embedded Linux kernel.
Engineering Challenges
Issue: Acoustic Reverberation in Enclosed Mechatronics Labs
Root Cause: Hard concrete and metallic test benches caused severe acoustic multipath reflection that corrupted uncalibrated microphone capture.
Resolution: Integrated adaptive spectral noise subtraction and beamformed spatial filtering across dual MEMS microphones, boosting Signal-to-Noise Ratio (SNR) by 14.8 dB.
Testing & Evaluation
Evaluated during rigorous live hardware demonstrations before the Ministerial Judging Committee at the National NURAI 2026 AI Competition, supplemented by automated offline test suites.
Results & Known Limitations
- ✓Awarded 'Good' evaluation rating by the Ministerial Judging Committee at the National NURAI 2026 AI Competition (Ministry of Higher Education and Scientific Research).
- ✓Demonstrated deterministic offline bilingual voice recognition and real-time mechatronic motor actuation.
- ✓Verified 100% operational uptime without an active Internet connection.
- ⚠Whisper INT8 acoustic model requires clear vocal delivery in environments where acoustic noise exceeds 85 dBA.
Future Improvements
- →Port acoustic matrix operations to an on-board FPGA coprocessor for sub-100ms multi-channel beamforming.
- →Integrate 6-DOF robotic manipulator inverse kinematics voice teleoperation.
Repository & Verification
Showcased at the National NURAI 2026 Competition · University of Basrah Team Zenith.