September Release Recap | AssemblyAI

This September, we shipped five major updates that make speech recognition more global, more accurate, and easier to implement than ever. Firstly, we ship the inapp playground, so you can test our speech to text capabilities directly from the browser with no code required. This makes testing our async and streaming models as easy as just a couple clicks. And you can use all the same features that are available via our API. and also provides the underlying code snippets so you can copy and paste the code directly to your IDE and get the same results enabling faster implementations. Secondly, we ship support for 99 languages for our universal 2 model up from the previously supported 17 with production grade accuracy and automatic language detection. Expanding your voice AI products globally is now easier than ever. Check out our supported languages page in the docs for more information such as what languages are supported and what speech understanding features are enabled for those languages. For the EU endpoint for EU data residency, we shipped PII audio reduction. If you want to redact audio files like this, — yes, last name is and my birthday is — just enable the redact PII audio parameter. You can specify policies to choose which entities you'd like to redact, such as people's names, phone numbers, credit card numbers, and many more. Switching gears now to our streaming model. We first shipped general improvements to the model itself, resulting in a 5% improvement to word error rate, 10% improvement to proper noun recognition, 41 millisecond improvement to latency, and much more. This requires no change to your code, and you automatically benefit from those improvements when using your streaming API. We also shipped key terms prompting, allowing you to specify a list of words you'd like to boost recognition accuracy for. For example, in this before and after, the model accurately recognizes both doctor's names and the medical terminology after providing it to the key terms prompting parameter. Our final update is that we added a new response field to our streaming API called utterance. The utterance field returns transcripts of segments of speech as soon as a short pause is detected. This was designed to return transcripts to you as fast as possible, typically within 250 milliseconds, so that you can start your LLM processing for your voice agent response as soon as possible, also known as preemptive generation. You can see in this side by side that the text is being returned much faster on the left, which is using the utterance field, rather than on the right, which is waiting for our turn detection model to determine the end of turn. Now, that's all for today. For the month of October, we have even more exciting updates to come, such as the universal 3 model and a streaming 2. 0 model, which will add multilingual support. If you'd like to try out our API today, check out our playground at assemblyai. com/playground. And we'll see you for next month for October's updates.