April 2023 Summaries

Word Error Rate (WER) is a commonly used metric for measuring the quality of speech recognition models, specifically automated speech recognition (ASR). It calculates the number of errors made by an ASR model in transcribing audio to text. The formula for WER involves counting the number of word insertions, deletions, and substitutions made by the model compared to a ground-truth transcript, then dividing this sum by the total number of words in the ground-truth. A lower WER indicates better performance. However, while WER is useful for comparing ASR models, it doesn't provide a comprehensive understanding of how well a model will perform on specific types of data or with certain vocabulary.

Deepgram, a remote-first company that grew by 9x during the pandemic, organized a week-long offsite in Cabo with three main goals: working together face-to-face, solidifying the shared understanding of the company's vision and goals, and enjoying each other's company. The event included team-based activities like "Dine-around Dinners," educational sessions, and karaoke night. Deepgrammers reported increased sense of belonging (94%), more clarity in the company's vision, mission, and strategy (92%), and greater confidence in leadership (81%). The offsite allowed employees to connect with each other, build a sense of belonging, and collaborate effectively.

OneReach.ai and Deepgram have partnered to integrate Deepgram's automatic speech recognition (ASR) technology into OneReach.ai's conversational AI platform, offering developers a low-code/no-code solution with over 10,000 features and a composable architecture for solving complex tasks. The partnership enables the creation of superior conversational solutions tailored to specific customer needs, faster deployment, and easier customization at a lower total cost of ownership (TCO). OneReach.ai has been recognized as a leader in Gartner's Magic Quadrant for Enterprise Conversational AI Platforms for two consecutive years, and the new Deepgram integration enhances its capabilities further.

The quality of data used in AI training is crucial for the effectiveness and reliability of AI models, particularly in high-stakes applications such as healthcare and public policy. Despite this, data work is often overlooked and undervalued, leading to "data cascades" that can have severe consequences on model performance. Addressing these issues requires prioritizing data quality, providing adequate training for data collectors, and recognizing the importance of domain expertise in high-stakes AI projects.

Obsidian is a powerful note-taking app that allows users to create a "second brain" for storing thoughts, plans, and general notes. One of its core plugins is the Audio Recorder plugin, which enables users to record audio for quick voice notes. To make these audio notes more actionable and searchable, Deepgram transcription has been added to the community plugin by Jason Maldonis. The integration allows users to transcribe their audio notes into text format within Obsidian. Setting up this feature involves installing the Audio Notes plugin, obtaining a Deepgram API Key, and configuring settings such as language, model tier, and keywords for context understanding. Once set up, users can record audio notes, save them with transcriptions, and edit or correct the transcriptions if needed.

Large Language Models (LLMs) are advanced AI systems that can understand human language and generate coherent text. They work through word vectorization, which transforms words into numerical lists for computation. Two popular training methods for LLMs include masked language models, where the AI fills in missing words in a sentence, and predictive language models, where the AI guesses the next word based on previous ones. The "large" aspect of LLMs refers to their number of parameters (at least 100 billion), which allows them to make accurate predictions about word placement in sentences. With these powerful tools, users can explore a wide range of creative applications and possibilities.

Deepgram introduces Starter Apps to make it easier for developers to build applications that transform real-time and prerecorded audio data into transcripts enriched with natural language understanding metadata. The apps are available in various languages, including JavaScript, Python, PHP, Go, C#, Java, and Ruby. Deepgram's Starter Apps provide a seamless onboarding experience for developers of all skill levels. Users can access popular models like Deepgram Nova and Whisper Cloud through these starter applications. To get started with the apps, users need to clone the repository, install with the language-specific package manager, edit the config file, and run the application.

Deepgram introduces Nova, a next-generation speech-to-text model that outperforms competitors in speed, accuracy, and cost. The fully managed Whisper API supports all five open-source models and offers built-in diarization, word-level timestamps, and an 80x higher file size limit compared to OpenAI's offering. Nova boasts a remarkable 22% reduction in word error rate (WER), blazing-fast inference time, and budget-friendly pricing starting at $0.0043/min. Deep

The Transformer model has significantly impacted natural language processing, influencing various subsequent models and techniques such as BERT, Transformer-XL, and RoBERTa. Its exceptional ability to understand and decipher the intricate structure of languages is due in part to its residual stream, which allows for effective communication between layers. Multi-head attention also plays a crucial role in the success of Transformers by enabling each head to work independently and contribute to more complex operations. Induction heads are specialized attention heads that enable pattern matching and remembering specific phrases or types of information. Overall, the versatility of Transformer-based models has led to their widespread use in various fields beyond natural language processing, including image processing, tabular data, recommendation systems, reinforcement learning, and generative learning.

This blog provides a step-by-step guide on how to transcribe audio files quickly using Google Colab and Deepgram. The process involves installing dependencies, uploading an audio file, transcribing the file, and viewing the transcription in JSON format. Users need to have a Deepgram account and API key to use this method. The blog also mentions that Deepgram supports various audio formats such as MP3, MP4, M4A, etc., making it versatile for different types of audio files.