1. Intelligent Audio Processing Solution
Customer Pain Points
In fields such as smart instruments, AI music creation, voice recording and monitoring, traditional solutions have problems such as high audio processing latency, poor real-time performance, and inaccurate audio feature extraction, making it difficult to meet the efficient processing needs of complex audio scenarios. For example, meeting transcription efficiency is low, and the misjudgment rate of sound monitoring in public places is high.
Solution Design
Using the Audio-M/X model of the RISC-V + self-developed DPU edge computing power development board, its DPU adds 1D FFT / MFCC dedicated instructions and SRAM double buffering, achieving streaming audio processing latency ≤2 ms; onboard 2-channel 24-bit ADC + 1 W Class-D DAC, supports multi-microphone circular array and beamforming technology, combined with built-in AEC/ANS algorithm library, can improve real-time voice enhancement SNR by 20 dB. With native support for PyTorch, TensorFlow frameworks, and reference models such as Whisper Tiny / Base, low-code integration can be quickly achieved through the Streaming ASR SDK for fast audio data collection, processing, and analysis.
Application Effect
Greatly improves the real-time performance and accuracy of audio processing, enabling real-time music creation and conversion in the field of smart instruments; in voice recording and monitoring scenarios, it can efficiently complete meeting transcription and environmental sound monitoring in public places, reduce misjudgment rate, and provide strong technical support for the intelligent audio industry.
2. Intelligent Vision Recognition Solution
Customer Pain Points
Scenarios such as access control gates, smart building attendance, campus perimeter security, and crowd behavior analysis have extremely high requirements for the speed, accuracy, and stability of vision recognition. Traditional solutions have problems such as slow recognition speed, insufficient multi-person scene processing capability, and low data comparison efficiency, which cannot meet the needs of large-scale personnel management and security prevention.
Solution Design
Using the Vision-F/G model development board, dual MIPI-CSI-2 and DPU collaboration, through hardware splicing of face/gait vectorization operators, builds a multi-model pipeline, achieving less than 18 ms processing time per 1080p image frame, supporting multi-person parallel recognition. With encrypted feature code offline comparison technology, more than 100,000 library comparisons can be completed within ≤0.3 s. Provides reference models such as YOLOv7-tiny-facial, ArcFace-50, combined with Python/C++ SDK + REST API, built-in liveness detection and personnel grading whitelist functions, ensuring the security and accuracy of recognition.
Application Effect
Significantly improves the efficiency and accuracy of intelligent vision recognition, enabling fast and contactless access in scenarios such as access control gates and building attendance; in campus perimeter security and crowd behavior analysis, it can monitor abnormal behavior in real time, effectively improving the level of security management, and providing reliable protection for smart campuses and smart building construction.
3. Live Streaming Edge Computing Solution
Customer Pain Points
In fields such as live e-commerce and real-time video creation, there are problems such as high video special effects processing latency, untimely intelligent subtitle generation, and low efficiency of barrage and traffic management, which affect the viewing experience and live broadcast effect, restricting the high-quality development of the live broadcast industry.
Solution Design
Using the Live-Edge model development board, H.264/H.265 1080p60 encoder/decoder is directly connected to the DPU to build a zero-copy inference path. Combined with image segmentation parallelism and ROI mechanism, the host face area is prioritized. Supports RTMP/SRT/WebRTC three-protocol aggregation, built-in CDN acquisition acceleration plugin to improve data transmission efficiency. With real-time human body segmentation (BodyPix), skin smoothing/filters, and other functions, as well as OBS Studio plugin, one-click acceleration channel can be enabled, achieving end-to-end latency of 80 ms for efficient live streaming processing on Windows/macOS/Linux systems.
Application Effect
Provides strong edge computing support for the live broadcast industry, enabling real-time face special effects, fast intelligent subtitle generation, and barrage/traffic QoS optimization, significantly improving live broadcast picture quality and audience interaction experience, helping live e-commerce improve sales conversion rate, and promoting innovation and development of the live broadcast industry.
5. Common Advantages and Supporting Services of the Solution
Common Advantages
The RISC-V + self-developed DPU edge computing power development board has equivalent computing power of 6 TOPS INT8 / 3 TOPS FP16, with typical power consumption of only 8 W, maintaining low energy consumption while performing high-performance computing; RISC-V 64-bit multi-core CPU and DPU collaboration reduces inference latency by more than 60% compared to pure CPU architecture. The AI development ecosystem is complete, providing ONNX & TFLite one-click quantization/compilation toolchain and a complete SDK, covering the entire process of model pruning, compilation, and deployment; supports Gigabit Ethernet, Wi-Fi 6, and other communication methods, with USB-C debugging and GPIO customization functions, adopts -20℃~70℃ industrial wide temperature design and aluminum-magnesium alloy heat dissipation base to ensure stable operation of the equipment; integrates hardware PMP, Secure Boot, supports RESTful API + MQTT/WebSocket and mainstream edge network management platforms, ensuring data security and convenient operation and maintenance.
Supporting Services
Provides rich development resources and professional services, including detailed technical documents and sample code, to help customers quickly get started with development; supports hardware customization and optimizes configuration according to customers' special needs; provides algorithm adaptation services to ensure efficient operation of models on the development board; assists customers in completing mass production testing and CE/FCC/CCC certification, providing customers with one-stop solutions to fully meet their business needs.
