September 2022 Summaries

This blog demonstrates how to build and use a multi-party video-conferencing application with LiveKit and Symbl.ai, two platforms that provide scalable, multi-user conferencing over WebRTC and out-of-the-box conversation intelligence capabilities. The application integrates with Symbl.ai's real-time APIs to provide features such as live closed captioning, real-time transcription, real-time insights, and sentiment analysis of conversations. The project was built using LiveKit's React SDK and Create React App, and the code is available in a GitHub repository. The multi-party video-conferencing application can be used for various use cases, including recruiting and HR intelligence, remote collaboration, revenue and sales intelligence, and trend analysis.

Symbl.ai's Managed Trackers Library is a collection of 40 pre-built trackers that can be used to gather actionable insights from conversations, making it easier for users to get started with conversation intelligence. The library provides well-researched and benchmarked trackers for various verticals such as Sales, Contact Center, and Recruitment. Symbl.ai's Tracker Recommendations feature proactively suggests new trackers based on the conversations processed using their Async API and Streaming API. This allows users to expand their set of trackers and continue getting new insights from their conversations. Custom trackers can also be created if managed trackers are not suitable for a particular use case, and Symbl.ai will find a match in the Managed Trackers Library and recommend it. The platform provides an easy-to-use interface to manage and customize tracked features, including Playground and API Explorer tools.

Conversations in customer interactions offer valuable data for organizations to gather feedback, improve products or services, and boost retention. Organizations need conversation intelligence tools to track and utilize these conversations effectively. Unlike documents, conversations require a different approach to analysis due to their free-flowing nature and contextual elements. Extracting value from human-to-human conversations can be done through real-time and asynchronous modes, which involve AI-powered actions such as sharing documents or offering discounts during calls, and analyzing data after the conversation ends. Advanced conversation intelligence has come close to helping employees better support customers mid-conversation by relating entities and temporally defining them, but it's a complex system that requires contextually understanding conversations. Organizations can leverage tools like Symbl.ai to process and analyze conversations across channels and extract value from customer conversations.

Audio classification is a process of categorizing audio files into categories according to shared features. It involves converting raw audio signals to digital representations, extracting relevant features such as amplitude, frequency, and time domain characteristics, and using machine learning algorithms to make predictions. The field has applications in various areas, including speech recognition, music classification, environmental sound detection, natural language processing, and noise-canceling, with deep learning approaches providing automatic feature extraction and improved accuracy.

The Context is Everything Hackathon at VOICE22, sponsored by Symbl.ai and Vonage, aims to bring creative project ideas to conversational insight, targeting applications where understanding conversation is crucial for evolution. The hybrid event allows anyone to participate without attending the conference, with a focus on using Symbl.ai's conversation intelligence platform in conjunction with Vonage's communication platforms. Cash prizes are awarded to encourage innovation, including $1,500 and a $500 Amazon gift card for first place, and additional prizes for specific categories or multimodal conversation analysis. The event is open to all, and projects must be submitted by October 10th to be considered.

Semi-supervised learning (SSL) combines labeled and unlabeled data to improve the accuracy of machine learning models, allowing them to learn from both scarce and abundant data sources. Unlike supervised learning, which relies on large amounts of labeled data, SSL uses a small portion of labeled data to train models and treats the rest as test data, reducing costs associated with labeling. In contrast to unsupervised learning, which does not require labeled data, SSL provides better learning results by using both labeled and unlabeled data. The technique offers benefits in areas such as speech recognition, audio classification, and web classification, where large amounts of unlabeled data are available but labeled data is scarce or expensive to acquire. By leveraging SSL, machine learning models can achieve improved accuracy and cost-effectiveness while handling challenges associated with missing labels in datasets.

The tutorial compares two popular machine learning (ML) and artificial intelligence (AI) frameworks, TensorFlow and PyTorch, by creating a feedforward neural network to classify handwritten digits from the MNIST dataset. Both frameworks are open source and free to use, enabling users to create, train, and deploy models for various ML and AI needs. The tutorial demonstrates how to architect the neural network using both frameworks, including the Sequential API and model subclassing approaches. It also shows how to train and evaluate the model's performance using common classification metrics such as precision, recall, and f1-score. The comparison highlights that TensorFlow is more beginner-friendly but PyTorch provides more flexibility for advanced or custom scenarios, leading to slightly worse performance in this particular case.

Symbl.ai is a conversational intelligence platform that offers abstractive summarization capabilities, supporting six dialects of English and processing audio, video, and text channels with no length limits. Its API provides pre-built requests in Postman for easy evaluation, and SDK support is available in JavaScript (NodeJS), C#, Python, and Web (TypeScript) for browser development. In contrast, GPT-3, a language model created by OpenAI, requires a third-party ASR service to process audio and video channels and has limitations on processing long-form media due to its 4,000 token limit. A side-by-side comparison of both platforms using a podcast video demonstrated that Symbl.ai can produce more detailed summaries than GPT-3, making it better suited for large-scale conversational intelligence efforts.

The text discusses how to enable developers to process contextual data in their applications using the Symbl platform's APIs, specifically for real-time and streaming applications. The post aims to provide a step-by-step guide on setting up a development environment with Node.js, installing necessary dependencies, and running example code to interact with the APIs. The authors highlight the simplicity of the Async API, which follows a standard pattern of login, process communication, check status, and obtain intelligence, while also emphasizing the flexibility of modifying existing examples to create more complex results and interactions between APIs. Additionally, the Streaming API is introduced as an even more direct approach, requiring only three steps: setting up JSON configuration, logging in, and attaching audio streams to WebSockets. The post concludes with a preview of future content, including a real-world project example that incorporates audio and/or video from streaming platforms.

This article discusses various types of machine learning classifiers, including their technical implementation, data types, techniques, and limitations. It also explores different technologies such as TensorFlow, PyTorch, Scikit-learn, and Armadillo for building classifiers, and highlights the importance of understanding the classification objective to limit the scope of a classifier. Additionally, it touches upon the use of Long Short Term Memory (LSTM) neural networks, Transformers, Conditional Generative Adversarial Networks, and Convolutional Recurrent Neural Networks in classifying text, audio, and video data. Furthermore, it introduces Symbl.ai as a conversational intelligence platform that simplifies classifying audio and video data by offering pre-trained machine learning models and APIs for easy integration with applications.