Make some noise for AI in JavaScript!
summary
Client-side noise suppression implemented for Infobip WebRTC uses an RNNoise-based model retrained with mixed clean speech and noise (MS-SNSD plus paid sets), runs on 40 ms frames with 20 ms overlap and produces ~24 gain outputs from 48 inputs, has under 1M parameters, and adds about 30 ms added latency (40 ms window, ~10 ms processing). Training used TensorFlow/Keras, MLflow and GPUs; inference runs in-browser via TensorFlow.js, numjs, AudioWorklet + Worker and successfully removes typing and static noise while sometimes over-suppressing music.
AI-written summary · about
chapters (10)
- 0:00AI noise suppression in JS
- 3:23About Infobip
- 4:48Real-time audio denoising problem
- 10:29AI noise suppression approaches
- 11:50Audio dataset preprocessing
- 13:41Neural network training
- 18:40WebRTC SDK implementation
- 21:33Live demos and results
- 26:20Challenges and future work
- 27:47Q&A
related talks
23mAudio and Video Manipulation
22mVoice Agents That Think While They Talk
39mMachine Learning Magic for Your JavaScript Application
01h 04m“Hello, AI?!” – Real-Time Voice Interactions with LLMs
01h 07mBuilding and Scaling an AI Agent Swarm of low latency real time voice bots
41mAI in Web: TensorFlow js
16mJavaScript Text Analytics for User Behavior Intelligence
30mLook ma, no hands! Multimodal AI-Agents in the browser