May 2023 Summaries

The article discusses the challenges of using non-differentiable loss functions in supervised learning models. It explains that while gradient descent is commonly used to optimize differentiable loss functions, it cannot be applied to non-differentiable ones. This limitation can be problematic for certain real-world applications where the loss function may not be differentiable. The author then explores alternative optimization techniques such as genetic and evolutionary algorithms, which do not rely on gradient information. However, these methods have their drawbacks, including slow training times and lack of GPU support. The article also introduces an adaptation of the Actor-Critic method from reinforcement learning to supervised learning. In this approach, a critic network is used to estimate the loss function's value, while an actor network updates its parameters based on the critic's output. This allows for training with non-differentiable loss functions and opens up possibilities for further exploration of alternative optimization techniques inspired by biological processes.

The development of AI voice synthesis has raised ethical questions about the use of deceased individuals' voices and potential misuse for fraudulent activities. Early text-to-speech systems used articulatory and concatenative synthesis, while more recent methods like statistical parametric speech synthesis have improved naturalness and cost-effectiveness. However, advancements in voice synthesis also bring challenges such as copyright issues, creator compensation, scams, security concerns for financial institutions, and potential misuse of evidence in trials. Efforts to detect synthetic voices are ongoing, but current methods may not be sufficient to keep up with the rapid advancement of AI voice synthesis technology.

Deepgram has introduced its fully managed Whisper API that supports all five open-source models and offers faster, more reliable, and cheaper services compared to OpenAI's offerings. The Deepgram Whisper Cloud is 3 times faster and 20% cheaper than OpenAI's Whisper Large model. It also includes built-in diarization, word-level timestamps, and an 80x higher file size limit. Deepgram has addressed the limitations of OpenAI's Whisper API by providing support for all five models, adding diarization and word-level timestamps, increasing the file size limit to 2 GB, and enabling on-prem deployments. The Deepgram Whisper "Large" model is priced at $0.0048/minute, making it more affordable than OpenAI's native offering.

This article provides a step-by-step guide on how to create speaker-labeled transcripts using Google Colab and Deepgram, an advanced speech recognition technology. The process involves installing necessary coding packages, uploading audio files into the notebook, modifying variables such as API key, MIMETYPE, and DIRECTORY, running the code to transcribe audio files, and finally creating a diarized transcript in human-readable format. Deepgram's capabilities also extend to summarizing or translating these transcripts.

Transformer models have made significant impacts beyond natural language processing (NLP), with applications in computer vision, tabular data modeling, recommender systems, and reinforcement learning. Vision Transformers (ViT) and DETR have emerged as strong competitors in image classification, object detection, and segmentation tasks. Transformers for tabular data include models like TabTransformer and SAINT that employ self-attention mechanisms to capture relationships between various vectors. In recommender systems, BERT4Rec uses the BERT architecture to build a powerful sequential recommendation system by capturing both short-term and long-term patterns. Transformers have also been adapted for reinforcement learning tasks, with models like Gated Transformer-XL and TrMRL showcasing improved performance and generalization capabilities. However, it's important to remember that self-attention is not the only solution for all machine learning problems, as evidenced by the need for hybrid architectures in some cases.

Deepgram has launched its API Playground, a tool designed to help developers, AI engineers, and data scientists quickly explore and test the Deepgram API in an isolated environment. The playground offers free unauthenticated usage for 60 minutes without sign-up requirements, support for all Deepgram language options, model options (Nova, Enhanced, Base & Whisper), and various features such as diarization, redaction, summarization, entity, and topic detection. Users can also access JSON explorer and transcription view of the response, along with code samples for cURL, node.js, python, and .Net. Deepgram offers $200 in credit upon signing up, allowing users to test the API without a credit card.

Ben Luks provides an accessible explanation of building a simple neural network from scratch using Python, aimed at demystifying AI for beginners. The project focuses on creating a basic linear regression model to predict the slope and bias of a line, using randomly initialized parameters, without relying on complex frameworks or terminology. The training process involves iteratively adjusting the parameters based on error calculations to improve prediction accuracy. Luks emphasizes the gradual learning process in machine learning, acknowledging the complexity of the underlying mathematics while encouraging readers to explore and learn progressively.

Deepgram has integrated with AudioCodes to power real-time voicebots and analytics, enhancing customer experience and operational efficiency in digital contact centers. The integration leverages Deepgram's industry-leading speech-to-text (STT) services within the AudioCodes VoiceAI Connect platform, enabling users to benefit from improved speed, accuracy, and ROI. This collaboration allows for seamless integration of voice bots into contact centers while supporting deployment on-premises as well as public and private clouds.

Recent research suggests that smaller language models (SLMs) are starting to outperform or match the performance of large language models (LLMs) in various applications, despite their larger counterparts' remarkable natural language understanding and generation capabilities. SLMs have several advantages over LLMs, including faster training and inference speeds, lower energy consumption, and reduced memory requirements. These efficiency benefits extend to other aspects of SLM use, such as smaller carbon and water footprints. As the focus shifts towards making AI more accessible and compatible with a broad range of devices, SLMs are becoming increasingly important in shaping the future of AI. Techniques like transfer learning, knowledge distillation, and specialized masking techniques have been employed to enhance the performance of SLMs. The potential for smaller models to achieve impressive performance gains without large-scale investment is showcased by recent techniques proposed by Google, UL2R, and Flan.

Language AI is transforming various industries and occupations globally. The technology includes Voice AI, Conversational AI, and Language Models like OpenAI's ChatGPT and Anthropic's Claude. Deep learning has played a significant role in the advancement of Language AI, enabling tasks such as speech-to-text transcription with human accuracy. A survey by Deepgram and Opus Research found that 82% of respondents use voice technology, and two-thirds believe it is crucial to their company's future strategy. The ability to customize Voice AI models is a major driver of value for businesses. Language AI has evolved from being a cost-saving tool to generating tangible value, with 79% of respondents reporting increased revenue after adopting speech technology. Companies that incorporate Language AI into their business processes will outcompete those that do not.

This article provides a step-by-step guide on how to transcribe YouTube videos programmatically. It covers the process of downloading audio from YouTube videos using the `youtube_dl` library and then transcribing the audio files using Deepgram's API. The code snippets provided can be used to transcribe multiple audio files at once, generating JSON files containing transcriptions with word-level timestamps. The article also suggests various ways to utilize these transcriptions, such as creating closed-captioning tools or translating videos into different languages.

Deepgram has released a new speaker diarization model that offers best-in-class accuracy and processes audio 10 times faster than its nearest competitor. The language-agnostic diarization model is free with all of the company's automatic speech recognition (ASR) models, including Nova and Whisper. Deepgram has also revamped its automatic language detection feature, resulting in a relative error rate improvement of up to 54.7% on high-demand languages such as English, Spanish, Hindi, and German. The company's large-scale multilingual training approach enables it to employ fast and lean networks while still obtaining world-class accuracy. Deepgram's diarization feature outperforms many commercial diarization models and common open-source alternatives like Pyannote when dealing with domain-specific, real-world data.

The article discusses the importance of using programming libraries like NumPy for efficiency and elegance in coding. It explains that while learning new frameworks or libraries may be burdensome, the benefits outweigh the costs. Using NumPy for simple calculations can result in faster outputs compared to manually written code due to its maximum parallelization capabilities. The article also suggests using other high-level libraries like PyTorch and Deepgram for specific tasks such as building neural networks or creating automated speech recognition models.

The article discusses Named Entity Recognition (NER) in recipes and how they can be treated as data. It introduces TASTEset, a dataset of 700 recipe ingredient lists annotated with named entities from nine classes. The author then explains the process of training a BERT-powered NER model for recipes using this dataset. They also discuss the results and inference of the trained model. The article emphasizes that even though the TASTEset code isn't entirely plug-and-play, it is still a valuable resource for those interested in parsing recipes en masse.

Rubber duck debugging is a powerful method for developers to work through coding issues by explaining the code as if talking to a rubber duck. This concept can be extended to other creative fields, such as writing, art, and design, with large language models (LLMs) serving as a "rubber duck" for creatives. LLMs are massive deep learning models trained on vast amounts of data that generate original outputs based on user prompts. These models have shown the ability to accelerate workflows in creative tasks by allowing users to quickly iterate ideas and engage in conversational dialogue with them, known as "prompt engineering." By serving as a sounding board for creative ideas, LLMs help bridge the gap between creativity and technology, making it easier for people to accomplish creative tasks.

OpenAI's new AI-powered chatbot, ChatGPT, has raised concerns about its tendency to output false yet plausible and coherent information. This phenomenon, known as hallucination, occurs when an AI model generates untruthful information on a closed-domain task. While hallucinations can lead to errors and offend in some cases, they also point towards AI that is not merely generative but potentially creative. Machine learning models are built around generalization, making the distinction between factual content and text structure, syntax, and delivery too fine for non-human models. Hallucinations have been used to generate MRI scans to supplement CT scans in lung tumor segmentation and improve autonomous navigation by hallucinating obstacles.

The article discusses the gap between AI fiction and reality, highlighting how pop culture has shaped our perception of artificial intelligence. It delves into recent developments in generative models like ChatGPT and Google's LaMDA, which have sparked conversations about their potential impact on society. Despite these advancements, the author emphasizes that current AI systems are far from sentient or autonomous, and warns against anthropomorphizing them. The article also explores how misleading media coverage can contribute to unrealistic expectations and hysteria surrounding AI technology.

TeleConnx and Deepgram have partnered to integrate Deepgram's real-time speech-to-text API into Asterisk-based PBX phone systems, offering a revolutionary solution for businesses managing phone calls. The new TeleConnx Speech-to-Text system enhances call routing in Asterisk and provides fast, accurate transcriptions through Deepgram's industry-leading technology. This integration helps organizations optimize their call handling process, automate routine tasks, and reduce response times, leading to improved customer satisfaction and valuable insights into customer behavior.

Large Language Models (LLMs) have shown remarkable capabilities in various natural language processing tasks such as summarization, translation, and generation. Recent advancements in LLMs have led to emergent abilities like text-to-code conversion, which has been utilized by products like Github's Copilot. As these models continue to scale in complexity and robustness, they are expected to yield more "second-order" applications beyond core NLP tasks. However, LLMs still suffer from hallucinations or non-factual responses due to erroneous encoding and decoding by the transformer or divergences in training data. To tackle these issues, researchers propose augmenting LLMs with external tools like calling another fine-tuned model, retrieving information via search engine or internet, solving computational problems via code interpreter or calculator, etc. This approach not only minimizes errors and hallucinations but also grants LLMs capabilities outside of textual generation. Several "second-order" applications have already been developed, such as intelligent software agents like Siri and Alexa, supercharged search engines, and hardware interfaces for robots. These applications leverage the model's response to perform another action or sequence of actions. As LLMs continue to evolve, it is possible that they may lead to even more abstract, complex "third-order" applications in the future, potentially bringing us closer to artificial general intelligence (AGI).

In this article, we learn how to create a YouTube video downloader using Python and Deepgram's API for transcribing audio. The process involves installing the necessary libraries such as pytube and requests, then writing code to extract only the audio from a YouTube video and send it to Deepgram for transcription. With Deepgram's summarization feature, users can also receive a summary of the video along with the transcript. This tool can be useful for saving time while consuming media or catching up on important content.