Our Products

Panda Agent Voiceprint Recognition Algorithm System

Voiceprint Recognition / Offline AI

Voiceprint RecognitionOffline AIEdge Deployment

Product Overview

Based on the WeSpeaker ResNet34 deep neural network model, it provides fully offline speaker enrollment and recognition capabilities. No internet connection required, supports local inference on edge devices, with recognition latency as low as 84ms (for 5-second speech). Suitable for scenarios with strict privacy and real-time requirements such as quadruped robots, industrial equipment, and smart access control.

Key Features

Speaker EnrollmentCapture ≥3s speech, extract 256-dimensional voiceprint feature vector and store in database
Speaker RecognitionCompare with voiceprint database, output best match and similarity score
Multi-user Voiceprint DatabaseSupports simultaneous enrollment of multiple users, verified for ≥4 concurrent recognition
Fully OfflineLocal ONNX inference, zero network dependency, data never leaves edge device
Edge DeploymentSupports Jetson Orin CPU/CUDA/TensorRT
Real-time ResponseRTF approx. 0.045, processing speed about 22 times real-time
USB MicrophoneSupports 16kHz mono audio capture, plug-and-play without drivers
Result AnnouncementRecognition results output in real-time via USB speaker or UART

Technical Specifications

Base ModelWeSpeaker ResNet34 CnCeleb (ONNX float32, 25.3MB)
Feature Dimension256-dimensional Speaker Embedding
Inference Latency5s audio CPU single-thread 227ms / 2-thread 134ms / 4-thread 91ms
RTFApprox. 0.045
Runtime EnvironmentPython 3.8+, sherpa-onnx, ONNX Runtime, Ubuntu 20.04

Typical Applications

Quadruped Robot Identity AuthenticationIndustrial Equipment Voiceprint Access ControlSmart SecurityVoice Command AuthorizationAttendance Verification