Speech Intelligence, Sentiment Analysis, and Auditory LLMs

From predicting the stock market to multimodal sentiment analysis, real-time intelligence is taking great strides. Here's the latest news!

Welcome (back) to AI Minds, a newsletter about the brainy and sometimes zany world of AI, brought to you by the Deepgram editorial team.

In this edition:

  • 🥽 A Deep Dive into Modern Sentiment Analysis

  • ⚒️ How to build an LLM Stack (Part 1)

  • 📈 Sentiment Analysis and Stock Market Prediction

  • 📸 Robust Multimodal Sentiment Analysis 

  • 📲 Trending AI Apps to Check Out!

  • 🎧 New AI Minds Podcast

  • 🎺 Demo: What fully AI-Generated Music Sounds Like

  • 🐦 Twitter: Financial Sentiment Analysis & Multimodal LLM Check-in

  • 📚 Algorithmic AI Glossary (code included!)

  • 🖥️ Upcoming Deepgram Webinar

Thanks for letting us crash your inbox; let’s party. 🎉

We coded with the brand-new Whisper-v3 over the past week, and the results were not what we expected. Check it out here!

🏇 Sentiment Analysis with Text and Audio!

Sentiment Analysis Deep-Dive: Teaching Machines about Emotions - In this article, ML Researcher Jose Francisco reveals why Sentiment Analysis is so difficult. After all, how hard could it be to classify text or audio as positive or negative? Oh, and the best part is that this article points out two ways for you to implement a sentiment analysis AI yourself!

Building an LLM Stack, Part 1: Implementing Encoders and Decoders - This page walks you through the first step of building an LLM from the ground up—from byte-pair encoding tokenization to grouped-query attention. If you’re hungry for an article that’s a bit more technical than usual, this guide is for you.

🧑‍🔬 Stock Market Prediction and Multimodal Sentiment Analysis

Sentiment analysis of financial Twitter - Researchers from Turkey use classic NLP architecture to scrape and intelligently analyze public sentiment towards a certain domain that could have influence on the stock market. Their results show over 10% increase in classification accuracy.

Modality translation-based multimodal sentiment analysis under uncertain missing modalities - This paper not only reveals the capabilities of current multimodal sentiment analysis models, but it also enhances them by coming up with a method of circumventing cases where not all modes of data are available for a given example. For instance, if a given datapoint comes with visual cues and audio, but no text, the “MJFs” within will compensate for the missing data.

WhisperTranscribe - Turn your audio and video into written content fast!

WhisperTranscribe makes it easy to get your audio transcribed in minutes. Once it's transcribed it works like your own ChatGPT but then for your audio. Create Summaries, Blog-posts, Show-notes or Social Media posts all written in your own style. '

Check this short demo out here

HeyDay makes another appearance, this time bringing to us their new product hunt launch. HeyDay is an AI copilot for knowledge workers that uses the insights in your research, conversations, notes and docs to assist you. 

Unlike ChatGPT, it integrates with all your sources of information and understands your knowledge to let you query your memory, create first drafts of content or be prepared for meetings.

🎙️ We are launching our new AI Minds Podcast this week!

The AI Minds newsletter has gained significant traction and is continually referenced as one of the best newsletters for AI news. We are now introducing a podcast aspect to the AI Minds brand.

Demetrios, Head of Community at Deepgram will be meeting with various founders in our start-up program and successful apps from our recently introduced AI apps page, where we will delve into their personal journey, company story and their learnings building an AI-enabled product. 

In our very first episode, Demetrios meets with Edward Brawer, CEO at PodcastAI, to discuss the value of AI in podcast production (quite apt for our first episode!)."

Also featured in upcoming episodes are, Samiur Rahman, CEO & Founder at Heyday, Nisha Iyer, Co-Founder at CoNote and Justin Jackson, CEO & Co-Founder at Transistor.fm and many more!

Each episode will be available in video and podcast format. Watch us on YouTube and also across all podcast platforms through Transistor.fm (Fun fact: they are also part of our start-up program and will feature in an AI Minds episode)

🎥 Stanford Lecture: Natural Language Understanding & Sentiment Analysis

In this Stanford Lecture, we learn about the challenges of Sentiment Analysis, gain some insight into its implementation, and learn ways to get started on building a sentiment analysis model itself. Start at 18:45 if you want to skip the contextualization and classroom logistics.

🐝 Social media buzz

📚 Glossary Pages

The glossary entries below outline everything you need to know about each respective topic—from histories, to use cases, to current implementations & resources. Check them out!

Sentiment Analysis: Of course, we had to include this glossary entry in this edition of AI Minds. Any relevant information that you think we missed above will surely be present here.

Bayesian Machine Learning: A couple of the sentiment analysis involved Bayesian inference among other Bayesian techniques and mathematical “tricks.” This glossary covers the essentials of such math.

Flajolet-Martin Algorithm: We’re including this glossary entry just to pique your curiosity. The Flajolet-Martin algorithm is a probabilistic tool that could not only navigate these vast data streams but also estimate their cardinality with remarkable efficiency

🖥️ Upcoming Webinar: An audio intelligence product showcase!

Tune in next week for the latest on Deepgram’s Audio Intelligence features and more product updates. Save your seat here!

Date: February 21st 2024

Time: 11am PT | 1pm CT | 2pm ET