Assembly AI TTS1 Targets More Natural Conversation (Full Transcript)

[00:00:05] Speaker 1: Hey everyone, this is Mark from Assembly. I'm here with Pascuale. He's the lead researcher on our latest text-to-speech model, Assembly AI TTS1. Pascuale, thanks for being here today. I read in the blog post there was this talk about naturalness in a text-to-speech model. Can you explain maybe like from a research perspective what naturalness means?

[00:00:26] Speaker 2: Yeah, I think our goal was really to try and capture natural conversation as it happens between individuals rather than focus on professional speech that you would get from scripted conversations.

[00:00:39] Speaker 1: So I guess in like the pacing, the tone, it's like two humans speaking to each other rather than someone reading something off the script?

[00:00:47] Speaker 2: Yeah, exactly. You want the model to speak to you as if it was a person at the end of your phone call, right? Not as a professional speaker reading out a book.

[00:00:57] Speaker 1: And how do you measure that like from a research perspective?

[00:01:01] Speaker 2: Yeah, that's a tricky question. So you could measure how close the prosody of your model is compared to data that comes from real conversations versus professional speech or like audiobooks.

[00:01:14] Speaker 1: We're all super excited about this launch. I just want to know like what's next for text-to-speech?

[00:01:19] Speaker 2: Yeah, we have a ton of features and updates we're working on. I think right now the highest priority is given to language expansion. We have a lot of new languages coming up and contextualization of the prosody of the model. So we want to add contextual information to the generation using information from previous turns to drive the prosody in a more natural way that is contextualized to the current conversation. Okay, thank you so much for your time. Thank you.

[00:01:46] Speaker 1: Bye-bye.

[00:01:47] Speaker 2: See ya. Bye. See ya.