Langtrace’s Post

To choose the optimal x264 preset, you must balance encoding and distribution costs. This article shows you how. The bottom line is that if your typical video gets over 500 or 1,000 views, it almost certainly makes sense to use the highest possible quality preset to get the bandwidth as low as possible. The Correct Way to Choose an x264 Preset: https://lttr.ai/AUSDZ #OptimalX264Preset

To choose the optimal x264 preset, you must balance encoding and distribution costs. This article shows you how. The bottom line is that if your typical video gets over 500 or 1,000 views, it almost certainly makes sense to use the highest possible quality preset to get the bandwidth as low as possible. The Correct Way to Choose an x264 Preset: https://lttr.ai/AUSDY #OptimalX264Preset

To choose the optimal x264 preset, you must balance encoding and distribution costs. This article shows you how. The bottom line is that if your typical video gets over 500 or 1,000 views, it almost certainly makes sense to use the highest possible quality preset to get the bandwidth as low as possible. The Correct Way to Choose an x264 Preset: https://lttr.ai/APUqu #OptimalX264Preset

🥧 DSPy is about optimizing that prompts, wherever they may fit into your pipeline, to get away from hard-coded prompt templates. Check out the full video: https://lnkd.in/gvSm4864

If you are interested about Multimodal LLMs (MLLMs) which can understand images and prompts on the image, their capabilities, and disadvantages, I invite you to read my new blog post :) Here, I analyze one particular MLLMs called mPLUG-Owl. I show how we can induce error to the responses of the model, but also to use the same model to find its errors! Yes, the model itself can be used to check if a previous answer was incorrect. Check my blog to know more, see you there. ✌

For any one working with text-to-speech (coqui TTS) and wondering how all the combinations of models + vocoder sound , have compiled them here for easy reference with some recommendations. With multiple models, datasets and vocoders to choose from, i wanted to get a baseline of which combination sounds better for my use case so that can serve as a base model for further fine tuning. Hence this compilation !! Hope it helps someone who is in similar journey or needs. Github link: https://lnkd.in/gjF5xNiy #tts #texttospeech #coqui

Feedback functions, analogous to labeling functions, provide a programmatic method for generating evaluations on an application run. It can be useful to think of the range of evaluations on two axis: Scalable and Meaningful. https://lnkd.in/eRQwfQsq

> Whisper Turbo < This is my second tutorial on Whisper and brings a few upgrades: 1. Turbo - 8X speed compared to Whisper large. 2. Free Colab notebook to transcribe any audio file (or YouTube Audio). 3. Consolidated notebook for dataset preparation, fine-tuning + conversion to OpenAI or CTranslate2 formats (for Faster Whisper). [Part of ADVANCED/transcription repo membership]. 4. One-click Faster Whisper server setup (incl. for custom fine-tuned models). Find it over on YouTube

Optimize LLM workflows with DSPy & Weave. Learn how to improve prompting strategies for causal reasoning using the BIG-Bench Hard benchmark. Explore our latest blog for a step-by-step guide. Read more: https://lnkd.in/gTQK8Zec

OpenAI has released a new optimized version of Whisper, 𝗪𝗵𝗶𝘀𝗽𝗲𝗿 𝗟𝗮𝗿𝗴𝗲-𝘃3 (𝗧𝘂𝗿𝗯𝗼), on GitHub. This new model has 809𝗠 𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿𝘀 and requires approximately 6𝗚𝗕 𝗼𝗳 𝗩𝗥𝗔𝗠. Here's a notebook for a quick comparison of the audio processing time between Whisper Large-v3 and Turbo on a 30-second audio clip. Notebook --> https://lnkd.in/gPeF9KBt