November 2022 Summaries

The Streaming Test Suite is an open-source project released by Deepgram to assist developers in getting started with their live streaming transcription API. It helps validate the functionality of the API, ensuring that the API key works, connection to Deepgram's API is successful, audio can be streamed to Deepgram, and transcriptions for the audio can be received. The test suite includes commented Python code demonstrating how to stream input from a microphone or WAV file to Deepgram and receive transcriptions back from their real-time endpoint. After successfully running the streaming test suite, developers will be ready to integrate Deepgram with more complex audio sources.

Developers can now use Deepgram's AI Speech Recognition Add-On for Twilio Programmable Voice, making it easier to unlock the value of post-call recordings. This integration allows developers to transcribe 100% of recorded calls and analyze them for sentiment analysis, topic trends, discussion points, and more. Deepgram's speech recognition technology offers high accuracy and various features such as punctuation, profanity filters, redaction, speaker diarization, named entity recognition (NER), multichannel support, numerals, callback, paragraphs, utterances, utterance split, and interim results. The integration is available in the Twilio console, where users can easily add Deepgram to their Programmable Voice project without needing a separate Deepgram account.

Deepgram built its platform in Rust due to the language's efficient memory management and strong concurrency support. The company had previously developed three iterations of their speech engine using Python, but faced challenges with memory allocation, bottlenecks, and parallelism. Rust provided a solution for these issues by offering better control over memory usage, enabling more efficient CPU utilization, and supporting multithreading without the risk of data races. Additionally, Rust's strict static analyzer, the borrow checker, helped reduce developers' cognitive burden by enforcing ownership invariants across complex programs. Deepgram transitioned to Rust for their platform, resulting in significant performance gains and improved product experience.

Deepgram has completed a $72 million Series B funding round, the largest raised by a speech AI company. The funds will be used to expand research and engineering teams, focusing on understanding and transforming speech at scale with end-to-end deep learning models. Deepgram's technology already offers fast, accurate, and cost-effective speech recognition, and the new features aim to provide more comprehensive insights into voice data, such as sentiment analysis, automatic summarization, speaker identification, and topic detection.

This article provides a step-by-step guide on how to migrate a React project to TypeScript. It begins by explaining the reasons for migration, including decreasing bugs, improving onboarding of new contributors, reducing the need for tests, making projects easier to maintain, and allowing quicker removal of technical debt. The author then outlines the process of getting started with the migration, adding TypeScript to an existing project, and finally migrating React files to TypeScript. The article also provides resources for learning more about TypeScript and accessing the code for this blog post.

Deepgram has released an Enhanced German (beta) Speech-to-Text Language Model, expanding its EU language support for the platform. The model is designed to accurately transcribe German speakers and can be used with various use case models such as Phone Call, Meeting, Voicemail, and Conversational AI. Developers can leverage this model in conjunction with speech understanding features like Diarization, Summarization, Topic Detection, and more. Deepgram plans to continue supporting the expansion of its EU language capabilities in the coming months to cater to its growing global customer base.

Deepgram has announced the availability of several on-prem Speech Understanding features including Summarization, Language Detection, and Topic Detection for its users. These features can be used with both Base and Enhanced models in an ASR request via the tier query parameter. Additionally, compatibility updates have been made to Deepgram Cloud and SDKs, allowing users to seamlessly transition between cloud and on-prem deployments. The latest release of Deepgram On-prem also supports half-precision floating-point format and the "CloseStream" web socket message for closing live audio streams.

The text discusses how Python and Deepgram's speech-to-text API can be used for compliance monitoring in call centers. It provides a step-by-step guide on how to set up the environment, install necessary packages, and write code that captures real-time audio streaming from microphones, transcribes it, and identifies each speaker in the conversation. The code also includes features like searching for specific phrases within the transcript and recognizing words not frequently encountered by the model. The author encourages readers to extend the project further using Deepgram's other streaming features.

This tutorial demonstrates how to create a workflow that transcribes voicemails and emails them using Twilio Studio, Pipedream, and Deepgram. The process involves creating a Twilio Studio Flow to handle incoming calls, prompting callers to leave a message, recording the caller's message, configuring a Twilio phone number to use this Flow, enabling the Deepgram add-on for Twilio, testing the webhook trigger, parsing the webhook payload, requesting transcription data from Twilio, and sending the transcript as an email.

This tutorial provides a step-by-step guide on how to transcribe podcast episode lists using Async IO with Python and speech-to-text providers like Deepgram. It explains the difference between synchronous and asynchronous programming, highlighting that while synchronous code runs tasks sequentially, asynchronous code allows for multiple tasks to be executed simultaneously. The tutorial then demonstrates how to use Deepgram's speech-to-text Python SDK to transcribe podcast audio files by creating a directory called "transcribe-audio-files" and writing the necessary Python code using async/await keywords.

The Learn Build Teach Hackathon, a collaboration between Discord and web developers, has announced its winners. Aadarsh Kannan's YouTXT project won the grand prize for using Deepgram's Speech-to-Text API to provide transcripts for YouTube videos. Runner up was DiscordSTT, a speech-to-text bot for Discord that enables users in voice channels to transcribe conversations. Both projects showcase innovative uses of Deepgram's technology and highlight the potential for improving accessibility and productivity.

Deepgram developed a custom Slack bot to improve customer support and productivity. The bot was built with the goal of providing best-in-class, personalized support at scale while maintaining the existing workflow of customers directly tagging support team members for attention. The development process involved gathering feedback from stakeholders, building the webhook handler, creating Zapier workflows, preparing for deployment, and testing before release. Since its implementation, the bot has allowed Deepgram to support more than 200 customers without significantly expanding headcount, resulting in faster response times and improved operational efficiency.

Deepgram has introduced new changes to its live streaming transcription API, making it easier to debug live stream requests. The platform now returns the request ID and any applicable error messages in the headers for the pre-WebSocket HTTP connection created for all streaming requests. This allows developers to view the HTTP headers of a failed handshake even if they cannot see the response body. An example using Python is provided, demonstrating how to use these new features. Deepgram plans to add more tooling over the coming months to further improve debugging capabilities.

This article discusses benchmarking OpenAI Whisper models for non-English automatic speech recognition (ASR). It covers the basics of measuring ASR model accuracy, challenges of accurate benchmarks for non-English languages, and benchmarking Whisper for Spanish, French, German, Hindi, and Turkish using curated publicly available data. The author highlights the importance of text normalization and consistent labels in ASR benchmarking and emphasizes that results should be contextualized by understanding the type of data the model is evaluated on.

Contributing to open-source projects can enhance a developer's resume by showcasing their ability to work on others' projects, communicate effectively with maintainers, adhere to guidelines, and demonstrate technical skills. It also improves the tech industry by ensuring software stability, growth, and security while reducing development costs. Open-source participation expands one's developer community, allowing for diverse perspectives and fostering a sense of belonging. Additionally, it provides a learning experience through exposure to different technologies, architectures, communication styles, and problem-solving approaches. Despite Hacktoberfest ending, developers are encouraged to continue contributing to open-source projects throughout the year.