Introducing Universal-3.6 Pro Realtime: Short Answers, Heard Right
Read the Portuguese translation →
Author: AssemblyAI · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Introducing Universal-3.6 Pro Realtime: Short Answers, Heard Right · Channel: AssemblyAI · 01:32 · Watch on YouTube
Topics: Model inference and optimization, AI models, Multimodal models
Summary
Section titled “Summary”AssemblyAI introduces Universal-3.6 Pro as its most powerful speech-to-text model for real-time use cases. The demo highlights real-time switching across four languages and support for 32 languages. The video presents improved performance in noisy environments and on one-word answers, including a word error rate as low as 1.45% for those utterances. It also describes keeping a call in its selected primary language when a caller switches languages, with accuracy improvements in Chinese, Arabic, and other languages.
Key points
Section titled “Key points”- Universal-3.6 Pro is presented as AssemblyAI’s most powerful speech-to-text model for real-time use cases.
- The demo shows real-time language switching and says the model supports 32 languages.
- The model is described as more accurate when background TV or other noise is loud.
- The video highlights improved handling of short, one-word answers.
- The word error rate for those utterances is reported as low as 1.45%.
- Calls can remain in the selected primary language even when the caller switches languages.
- Accuracy improvements are noted for Chinese, Arabic, and other languages.
- The model is available through the AssemblyAI API.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Introducing Universal-3.6 Pro | AssemblyAI introduces its real-time speech-to-text model and demonstrates language switching, noting support for 32 languages. |
| 00:30 | Performance in noise and short answers | The video describes improved accuracy with loud background noise and better recognition of one-word responses. |
| 01:00 | Staying in the caller’s language | The model is described as maintaining the selected primary language across language switches, with accuracy gains in several languages, before the API call to action. |
Quotes
Section titled “Quotes”our most powerful speech-to-text model for real-time use cases — 00:00
now 32 of them — 00:30
our word error rate has gone down as low as 1.45% — 01:00
how you use Universal 3.6 Pro through our API. — 01:32
Concepts and entities
Section titled “Concepts and entities”Universal-3.6 Pro, AssemblyAI, speech-to-text, word error rate, API
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Today, we are releasing our most powerful speech-to-text model for real-time use cases >> [music] >> yet. Let me show you with a quick demo. I’m going to first talk in English. What we show Indian John Wayne May Hindi be able to speak to you? I have to speak in your own poquito poquito. You see four languages switched real-time with the lowest latency and we not just support four languages, but now 32 of them. Now, here are the ways our
00:30 model performs better than ever before. Number one, during heavy noises. Think about a scenario when you have TV playing in the background or there’s heavy noise. Here’s an example. >> [music] >> full chamber next >> send it to my card and stop? >> Meteorologists expect heavy rain across the north >> [music] >> And the accuracy is proven to be way higher when the TV or the noise is louder. Number two, during one-word answers. For example, let’s say an agent asks you, “Can I have [music] a have this meeting at 3:00 p.m.?” And now when you respond
01:00 >> Sure. Stocks closed higher in the >> [music] >> During those utterances, our word error rate has gone down as low as 1.45% and number three, staying in the caller’s language. Let’s say you do a call in English and slowly sometimes switch to Spanish, but the entire steering of that call will remain in the primary language that you have chosen. And here you can see noticeable accuracy improvement in Chinese, Arabic, and other languages and now we support 32 languages. And now we cannot wait to see
01:32 how you use Universal 3.6 Pro through our API. Thank you.