Inspiration

Technical interviews, presentations, salary negotiations, sharing bad news, thesis defenses … what do all these have in common?

You need to practice. You need feedback.

TalkThru is the professional you need for on-demand, face-to-face practice. Technical or behavioral, emotional or cold: test your speeches, knowledge & improvisation with immediate corrections.

What it does

You provide a topic to the agent. You provide the agent a role.

The agent holds a conversation with you and summarizes what you should improve on at the end. It’s that easy.

How we built it

5 steps:

Program initialization: User provides role & theme for the interview.
Query: a) Share the prompt with Perplexity to generate an introduction, b) generate a fitting video with LumaLabs.
ElevenLabs: translate Perplexity's text-based response to speech & begin the conversation.
Speech2Text: The user's response to the agent's question is translated to text (using python speech_recognition).
Re-query: Unless the user wants to end the conversation, repeat steps 2-4.

Challenges we ran into

Speech-to-text: pause identification, fast responses, storing & processing audio real-time were challenging aspects of the speech integration.
Lip-syncing: realistic interview preparation requires mouth movement. Whereas LumaLabs provides vastly creative features, it does not support facial expressions. Our first feature release will enable lip syncing so that users can carry realistic conversations.
CUDA integration: lip-syncing requires torch access, CUDA-based GPUs & many pre-trained models. Thanks to our sponsor NVIDIA, we were able to develop this feature.