Product 1: Audio-M/X

1. Usage scenarios

Smart musical instruments, AI music.

Call center, transcription of meeting minutes, ambient sound monitoring in public places, wearable voice assistant.

2. Professional optimization

· 1D FFT/MFCC dedicated instructions & SRAM double-buffered in the DPU for ≤2 ms latency for streaming audio processing.

Onboard 2-channel 24-bit ADC+1W Class-D DAC, supporting multi-microphone ring array/beamforming.

Built-in AEC/ANS algorithm library, real-time voice enhancement SNR increased by 20 dB.

3. Out-of-the-box capability

Reference Model:

-WhisperTiny/Base On-Device inference 150 ms/30s fragment.

-PianoSession piano music real-time midi stream conversion.

-AudioLM Noise Detection, BirdNet Offline Bioacoustic Detection.

· Streaming ASR SDK: WebSocket/GRPC and low-code integration.

Product 2: Vision-F/G

1. Usage scenarios

Analysis of access control gates, intelligent building attendance, park perimeter, and crowd behavior.

2. Professional optimization

Dual-channel MIPI-CSI-2+DPU for face/gait vectorization operator (CosFace/GaitSet) hardware splicing.

Multi-model pipeline: detects one-alignment and one-feature extraction, 1080p<\18 ms per frame, and supports multi-person parallelism.

Encryption feature code (256 B) offline comparison > 100,000 databases, ≤ 0.3s.

3. Out-of-the-box capability

Reference models: YOLOv7-tiny-facial, ArcFace-50, GaitGL-base, with an official quantization accuracy loss of ≤0.5%.

Provide Python/C++SDK+RESTAPI, built-in liveness detection, personnel classification and whitelist.

Product 3: Live-Edge

1. Usage scenarios

Live e-commerce camera insertion, real-time face effects, sound beautification, intelligent subtitles, barrage/traffic QoS optimization.

2. Professional optimization

· The H.264/H.265 1080060 codec is directly connected to the DPU, and the inference path is zero.

Image fragmentation parallel + ROl (Region of Interest) mechanism, which prioritizes the facial area of the anchor.

· RTMP/SRT/WebRTC three-way protocol aggregation with built-in CDN collection acceleration plug-in.

3. Out-of-the-box capability

Real-time body segmentation (BodyPix), dermabrasion/filters, speech recognition and subtitles with 80 ms end-to-end latency.

· OBS Studio plug-in: enable the Live-Edge acceleration channel with one click, supporting Windows, macOS, and Linux.