Nishant Nikhil
@nishnik
Followers
1K
Following
9K
Media
83
Statuses
994
ML Research @meta | Text to Speech @play_ht | previously founded https://t.co/4WsLeE45X6 | Ex-Amazon ML Applied Scientist | Where is the music?
Bengaluru, India
Joined May 2011
Come see our work on "Training Language Models under Resource Constraints for Adversarial Advertisement Detection" today at 8:10 AM IST (19:40 PDT). We explore weak supervision, curriculum learning and multi-lingual training. PDF: @NAACLHLT #naacl2021.
2
1
32
Stoked to join @Meta Superintelligence Lab, and excited to train better models at 🚀 speed!.
Congratulations to @_mfelfel @HammadH4 @keikumata and the entire @PlayAIOfficial team on their acquisition by @Meta Superintelligence Lab! . All of us @kindredventures are thankful and honored to be part of their journey as their lead seed investor last year. I'm stoked for.
15
3
171
🚀 Just dropped our first research blog! 🎉. We trained a discrete-diffusion model (think Gemini Diffusion ✨) that handles TTS 🎙️, adds words ➕, and even removes words ❌ from speech! Open-sourced 🌐. Check out our blog 📝, GitHub 🐙, and Hugging Face 🤗 links in the comments!.
We’ve written more about how the model works, its architecture, and training setup in our blog. If you're building in this space or want to contribute, we’d love to connect. More here:
2
0
10
Had a great time chatting with .@OrkoHunter on his podcast. Always fun to explore ideas out loud. :). Hope you enjoy listening as much as I enjoyed the conversation.
🚨 New Podcast Alert!. Just dropped a fresh episode of The OrkoHunter Podcast with my good friend and applied AI genius @nishnik !. Topics.🤖 AI & the fascinating history of Deep Learning.🧠 Why Large Language Models matter today.🔥 What the future of tech might hold (and yes,
0
0
13
RT @OrkoHunter: 🚨 New Podcast Alert!. Just dropped a fresh episode of The OrkoHunter Podcast with my good friend and applied AI genius @nis….
0
1
0
As natural sounding as it gets, we have built an end-to-end speech model. That means if you are talking with this model, it takes into consideration how you are speaking, the emotions conveyed; and then articulates a reply with empathy :) ✈️🤖.
Today, we're introducing PlayDialog beta. It's a SOTA voice model built for fluid, emotive, human-like conversation. PlayDialog lets you create engaging dialog and speaker narrative programmatically -- and from documents, URLs, or text. Nothing else sounds as human.
5
4
49
With a sub-200ms time to first byte, we’d love to hear about the voice agents you’re building!.
Today we’re introducing our latest Text-To-Speech model, Play 3.0 mini. It’s faster, more accurate, handles multiple languages, supports streaming from LLMs, and it’s more cost-efficient than ever before. Try it out here:.
0
0
4
RT @arpit_tarang: Here's what @ppsreejith_ @nuwandavek and I have been up to for the last few months! Please check it out:.
0
3
0
random reminder about pair programming.
newyorker.com
Coding together at the same computer, Jeff Dean and Sanjay Ghemawat changed the course of the company—and the Internet.
1
1
11
Folks are going to build distribution on GPT-4/Llama prompting, and then come back to training their own model.
0
0
4
RT @arankomatsuzaki: Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models. Substantially improving over the….
0
56
0
added a website. (still on github pages) - vv static. Going to start a t2? and put it behind.
nishnik.com
Nishant's writings on deep learning, natural language processing and computer vision.
0
0
1
> Expected it. > tokens are registers, registers tokens. > vvv simple. > not filler tokens, <pause> tokens. That blue bar below is having registers (they call it pause tokens)
1
0
2
Vision models benefit from registers. The sink token looks like a register. We will see more registers coming to the language models.
0
0
2