Hey guys! So glad to be in here. I have been following nour's youtube videos for a while and am here to learn and provide value. Just wanted to share something thats helped me improve my voice AI pipeline a lot! I spent a lot of time trying to optimize my pipeline by swapping models. While that wasnt the best fix i found what worked was measuring each stage separately first - STT, LLM, TTS, and network - before touching anything. mine turned out to be a stage I'd never suspected. Also surprisingly logging timestamps at every stage boundary and looking at real numbers . And streaming everything , so not waiting for full utterance to hand off to the next stage. Most of my "model latency" was just waiting After lots of testing, my current stack sits around 290ms end to end.. Curious what everyone else is measuring - and where your bottleneck actually turned out to be?

I didn't write this, and have not tried it. But I thought this something the community might be interested in. https://medium.com/data-science-collective/build-a-real-time-voice-agent-in-30-minutes-with-interruption-handling-fa67a926b5f9

Hi Folks, I'm planning to self-host LiveKit Egress in Kubernetes and would like some guidance on infrastructure sizing and autoscaling. Our use case - Track Egress (not RoomComposite) - Supporting both 1:1 video calls and group video calls - Recording each participant's audio and video tracks separately - Maximum recording resolution: 720p (1280×720) - Maximum frame rate: 30 FPS - A separate merge service combines the recorded tracks into a final MP4 - Final recordings are uploaded to S3 - We are planning for hundreds of concurrent recordings Questions 1. Is there an official recommendation for the minimum and recommended CPU/RAM required per Egress instance for Track Egress at 720p/30 FPS? 2. Approximately how many concurrent Track Egress jobs can we expect from servers with: - 4 vCPU / 8 GB RAM - 8 vCPU / 16 GB RAM - 16 vCPU / 32 GB RAM - For 1:1 calls, if both participants have audio and video tracks, are there any specific resource considerations compared with recording a single track? - For group calls, where multiple participants' audio/video tracks are being recorded separately, how does the number of tracks per room affect Egress CPU, memory, and network usage? - For a maximum of 720p/30 FPS, what factors have the biggest impact on CPU and memory consumption for Track Egress? - Are there recommended Kubernetes CPU/memory requests and limits for Track Egress? - What is the recommended approach for autoscaling Egress workers in Kubernetes? Is CPU-based HPA sufficient, or is there a better metric/approach for determining when another Egress worker is required? - Since we expect hundreds of concurrent recordings, is there a recommended architecture where LiveKit Egress is hosted separately from the LiveKit Server and scaled independently? - Is there any managed/cloud platform or LiveKit-supported service available where Egress workers can be hosted separately with automatic scaling, rather than managing the Egress Kubernetes deployment ourselves? - Are there any production examples or recommendations for sizing a large-scale Track Egress deployment supporting both 1:1 and group calls?

Hi, i have tried Assembly ai, and Deepgram's latest models (flux) as they are promising accurate turn handling (with barge in functionality) also. However on testing, neither is reliable...the accuracy of transcripts is lower than Nova models ...does anyone know how to solve this. ?? any models or settings which is accurate ? Also has anyone used the pipecat smart turn detection? Does is work well to indentify pauses and thinking vs end of sentence like the semantic based turn detection in flux/ assembly ai .?

Hello everyone. My name is Umit. I am 64 years old and facing a few health challenges, most notably memory issues. I aim to keep my brain active and slow down their progression by joining AI communities like Open Source Voice AI community. My purpose in joining is not for financial gain at all; the more I focus on learning, the better I feel. Thank you for welcoming me into your community.