Translate tech talks using open-source models and cloud-native Wasm
summary
VideoLangua runs a concrete pipeline: Whisper transcribes, Llama 3.18B chunks and Q1/Qwen-style models (Q1 2.57B or Qwen-72B depending on target) translate, a small Llama 3.2–3B cross-checks, and GPT-SoVITS or TTS renders voice, with human review optional. Runtime is a tiny Wasm API server (API server ~3 MB, models larger) on WasmEdge / wasmagic enabling portability across spot GPUs and lower cost (uploader cites $80,000/day vs ~$100/day). Highlights: agent collaboration across models, real-time vs high-accuracy inference tradeoff, Wasm portability for heterogeneous GPUs.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:01Live demo: uVideoLangua
- 10:00Agent collaboration for accuracy
- 15:05Why WebAssembly powers translation
- 22:41Running Whisper locally
- 26:37Enterprise use case: Clapper
- 30:23Q&A
related talks
40mOpen Source LLMs in the Cloud: Scalable Solutions
48mWorkshop: Efficient and Portable AI / LLM Inference on the Edge Cloud
31mDeploying Large Language Models with Ease
25mLeveraging Wasm for Portable AI Inference Across GPUs, CPUs, OS & Cloud-Nativ...
10mLightning Talk: Scaling AI With Wasm and Edge Computing
31mFrom ML to LLM on device AI in the browser
27mWasm Is Becoming the Runtime for LLMs
30mOrchestrating Real-Time Multimodal AI Agents with Rust