April 2021 Summaries

The AssemblyAI Speech-to-Text API now offers two new endpoints that allow developers to pull a completed transcript broken into paragraphs or sentences, making it easier for users to digest the information. To access this feature, no additional account setup is required, and existing audio files can be used without resubmission. The new endpoints are /v2/transcript/{TRANSCRIPT ID}/paragraphs and /v2/transcript/{TRANSCRIPT ID}/sentences. An example of how to use these endpoints in Python is provided along with a sample response from the API. This update makes it simpler for developers to create user-friendly interfaces directly from audio files.

AssemblyAI uses Drone as their primary CI/CD tool, but needed a cost-effective and easily scalable solution for their GPU instances. They decided to use Docker Runners and built drone-queue-cloudwatch, an open-source tool that sends Drone build stats to Cloudwatch and lets it handle scaling. The tool uses Node Routing to define which worker group handles a given build and publishes metrics for running builds in order to control scaling. It also considers the build queue depth as the effective metric for scaling. The code is available on GitHub, and there are example Terraform files for implementation.

HttpClientFactory is a useful tool in .NET ecosystem that provides an opinionated factory for creating HttpClient instances. It addresses issues related to managing the lifetime of HttpClientHandlers and helps avoid problems with trying to handle HttpClient lifetimes manually. In this post, three basic implementations of HttpClientFactory are discussed: using it directly, named clients, and typed clients. Each implementation offers different benefits in terms of configuration, reusability, and testability.

The internet is flooded with audio and video content, much of which may contain abusive language or excessive profanity. To moderate this content, large teams of people are required to manually review it. However, AssemblyAI has introduced a new feature called Content Safety Detection that automatically classifies transcriptions into over 17 labels such as "hate speech", "profanity", and "NSFW". This feature is powered by state-of-the-art deep learning research and can be easily integrated into any audio or video processing pipeline. It helps developers find sensitive content in their files without the need for human intervention. The JSON response from the API includes specific sections of a transcription that are flagged, along with timestamps and confidence scores for each detected label. This feature is now available globally to all developers.