Nikita Torgashov’s Post

🎬 VP9 SVC enables our Media Router to deliver different video resolutions from a single encoded stream, providing many benefits like better video quality on low-bandwidth connections, or improved network efficiency with reduced latency. 💡This technology is more efficient than simulcast and works seamlessly across Chrome, Safari, and Edge. 👀Ready to upgrade your video experience? Read our full blog post for implementation details and technical insights -> https://gag.gl/v8jQTo #mycompany

🎬 VP9 SVC enables our Media Router to deliver different video resolutions from a single encoded stream, providing many benefits like better video quality on low-bandwidth connections, or improved network efficiency with reduced latency. 💡This technology is more efficient than simulcast and works seamlessly across Chrome, Safari, and Edge. 👀Ready to upgrade your video experience? Read our full blog post for implementation details and technical insights -> https://gag.gl/v8jQTo #mycompany

“During testing, we’ve deployed AV1-encoded video to randomly selected groups of users. The feedback is significantly higher for AV1 versus other codecs.” ~ Meta Research Scientist and Technical Leader Ioannis Katsavounidis  Read more in Meta's AV1 Adoption Story 📚 https://brnw.ch/21wWkgL

🎧 System Design Interview Question (Google) Q: Design a denoising system for audio that supports both real-time and batch applications — balancing latency, quality, and deployability. A (sneak peek): Hybrid DSP + ML approach Real-time path: RNNoise-style low-latency models Batch path: Demucs / Conv-TasNet for high-quality post-processing gRPC streaming APIs for live audio GPU-backed inference fleets and autoscaling Quality metrics: SNR, PESQ, and human MOS Full breakdown 👉 https://lnkd.in/ec3i_2W9

How fast is your speech-to-text really? Latency isn’t one number—it’s a timeline. In our new deep dive, we share a reproducible benchmark and practical SLOs for real-time apps. What's inside: • A practical overview of key latency metrics and definitions: TTFB, partials vs. finals, endpointing latency, RTF • Streaming tips: frame size, network/jitter profiling, and balancing trade-offs (speed vs. accuracy) • Comprehensive guidelines and concrete SLOs you can apply when evaluating STT latency for your application 👉 Read the article: https://lnkd.in/enW22Kri

Sometimes the next billion-dollar company starts as a doodle (not the dog btw). At PWV, we believe some of the biggest ideas in don’t always begin with a pitch deck: they begin with curiosity, creativity, and most likely some experimental lines of code. That’s how we saw tldraw’s playful “Draw Fast” demo evolve into fal's generative media infrastructure, now a $1.5B company powering over 1M developers. Our latest post from GP, David Thyresson explores how small experiments can reveal massive potential, and why “seeing the big picture” often starts with seeing what others overlook.

𝗘𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘀𝗮𝗶𝗱 𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗶𝗻𝗴 𝗔𝘀𝘁𝗲𝗿𝗶𝘀𝗸 𝘁𝗼 Pipecat 𝗰𝗼𝘂𝗹𝗱𝗻’𝘁 𝗯𝗲 𝗱𝗼𝗻𝗲. 𝗜 𝗽𝗿𝗼𝘃𝗲𝗱 𝘁𝗵𝗲𝗺 𝘄𝗿𝗼𝗻𝗴 𝘂𝘀𝗶𝗻𝗴 𝗚𝗼. The problem? There’s no defined or reliable way to transmit real-time voice from Asterisk’s AudioSocket into Pipecat. Even after downsampling or upsampling audio to the proper rate, custom serializers, service validation, and transport adjustments, it simply didn’t work. No audio at all. Just silence and endless debugging. So I built a Go-based AudioSocket implementation that finally made them talk. It handles real-time audio, DTMF events, and hangup (SLIN), plus custom 𝗣𝗶𝗽𝗲𝗰𝗮𝘁 𝘀𝗲𝗿𝗶𝗮𝗹𝗶𝘇𝗲𝗿𝘀, all using existing 𝗪𝗲𝗯𝗦𝗼𝗰𝗸𝗲𝘁 𝘁𝗿𝗮𝗻𝘀𝗽𝗼𝗿𝘁. No need for a custom layer in pipecat itself, and since it’s Go, it can scale effortlessly to hundreds of concurrent connections without breaking a sweat. I’ll be releasing a full video breakdown soon. If your team is wrestling with fragile real-time audio or AI voice integrations, that’s exactly the kind of rescue and production hardening we specialize in, turning messy prototypes into reliable, scalable systems. #Asterisk #Golang #Pipecat #VoiceAI #Telephony #RealTime #WebRTC #Concurrency #OpenSource

For any of those interested in open source video generation. Best read up. Wan 2.5 preview is behind closed doors for now, and perhaps will not be released, but my sources indicate that Wan 2.5 will eventually be released and open source. Given the communities ability to quantize models, once Wan 2.5 is released, it will quickly get quantized and be able to run on fairly modest consumer GPUs. https://lnkd.in/g3g_37Kk

Sora 2 can do things that are exceptionally difficult for prior video generation models. It's more physically accurate and realistic than prior systems and comes with synchronized audio. It's impressive to see how OpenAI, which had been lagging behind in video generation, has finally managed to solve the physics of gymnasts where other models still struggle. Let's see how long it will take open source to catch up this time!

the ttfb (time to first byte) metric for measuring STT latency for voice agents is inaccurate, here's why you should be using ttct (time to complete transcript) instead: ttfb only measures silence -> start of speech -> first transcript chunk it completely ignores the more critical path which is ttct: end of speech -> silence -> final complete transcript (to be sent to the LLM) this can lead to some interesting conclusions: - streaming models that emit fast but inaccurate partials get rewarded (easy to game) - async models ran in a low latency fashion look like they perform terrible with ttfb due to no mid-speech partials (even tho they can work well in a voice agent) but the thing is: you're not using partials to generate the LLM response, you're using finals, so that is what you should measure models also have different configs for emitting partials vs finals, with partials you want to optimise for speed, finals you want to optimise for accuracy (without affecting latency too much) - so you will get different speeds at AssemblyAI we're laser focused on optimising for TTCT while maintaining industry leading accuracy, and we're shipping a new update today to our streaming model to return transcripts faster than any other provider via a new field called "Utterance" which emits finals on any 160ms silence so you can generate your LLM response as quick as possible - give your voice agents a noticeable speed up today by testing it out ;)