MARS5 TTS

@akshat_prakash2 Hi CAMB.AI Team, Just saw MARS5 on Product Hunt and the focus on open-source, prosody replication, and multilingual capabilities is exciting! It sounds like a valuable tool for developers and creatives seeking more expressive text-to-speech solutions. A couple of questions piqued my interest: - How does MARS5's performance (audio quality, naturalness) compare to existing open-source and closed-source TTS models, particularly for challenging prosody? Are there plans for easy integration with existing development tools? - Does MARS5 offer any user control over the prosodic aspects of the generated speech? What are your plans for future development, such as expanding language support or offering different "voice styles"? My background in AI development might be valuable for MARS5's development in areas like: - Exploring ways to further enhance MARS5's audio quality, naturalness, and robustness for various prosodic scenarios. - Developing user-friendly tools and libraries that streamline integration of MARS5 into various development workflows. - Investigating ways for users to customize speech prosody and potentially offer different voice styles or emotional variations. I'd love to learn more about MARS5's technical details and roadmap. I'm eager to see how my skills can contribute to making it the leading open-source TTS solution for replicating diverse prosody across languages. Thanks for launching such an innovative project! Best regards, Max

@max_savonin1 Hey Max, Thanks for the questions and the interest shown! - MARS5 is a significant step ahead of most opensource and closed source TTS models in its ability to capture prosody for tough scenarios like sports, anime, movies etc. MARS is already integrated into CAMB.AI's paid offerings such as -- DubStudio (studio.camb.ai), APIs (docs.camb.ai) and our chrome extension (https://chromewebstore.google.co...) - MARS5 replicates the reference speech given, so yes, it allows for controlling prosody in that manner. Since it's auto-regressive, you can generate a few times until you find one that's best, and we're baking in more inference techniques for higher quality and stability very quickly into the release. - If you're interested in contributing and joining the team, definitely do make atleast 1 pull request into our repo and feel free then to email us. We're hiring!

@akshat_prakash2 looks impressive. Would love to see more of the interface and features before having to sign up. Do you envision implementing this at the platform level for video sites? More thoughts

@frank_denbow hey Frank, I see you're a fellow Carnegie Mellon homeboi; thanks so much for the review. I'm class of SCS as well. GO TARTANS!!!

@akshat_prakash2 Looks fly! Would love to connect!

@charlotte_hare do join our Discord and let us know how we can help. Thanks for trying it out!

@muhammadnoval1 Thank you noval!!!

@sentry_co Thank you!! I think it's multi-linguality and stability for sure.

@njrobb hey thanks for asking -- it's charged per minute of output!

@akshat_prakash2 Thanks for getting back. At what rate?

@guillaumemdsm thanks, I'll keep a look out for a pull request!

@hugec thanks!