Taughtful.ai: a new way to learn, powered by Vocal Bridge

A live tutor for a document, video, course, or topic that teaches out loud, draws fresh slides, and changes direction when the learner interrupts.

Why passive lessons fall short

Recorded video and static documents move in one direction. They can present material, but they cannot notice the moment a learner gets stuck. Research by Guo, Kim, and Rubin, based on 6.9 million video-watching sessions across four edX courses, found that median engagement time was at most six minutes regardless of video length (ACM Learning@Scale, 2014).

Taughtful.ai starts from a different premise. Learning is a loop between an explanation and the person trying to understand it. That loop works best when the learner can say “wait,” ask for a sports analogy, switch languages, or go deeper while the lesson is still unfolding. The next lesson can then begin with a better sense of how that learner prefers to be taught.

Sapna Sangmitra describes the longer-term version of that loop:

"My favorite part is the post-processing one. What it does is whenever you go through the lesson, it creates a transcription of it and goes through it to understand how you are learning, the way you like to learn."
- Sapna Sangmitra, Founder of Taughtful.ai

What Taughtful.ai built

Taughtful is a spoken tutor rather than a summary tool. A learner can bring a public YouTube video, upload documents, combine sources, name a topic, or open a course. Taughtful indexes the material into sections and turns it into a guided lesson with a live teacher agent.

The tutor does not speak over a static page. Taughtful's Generative Player composes visual explanations as the lesson moves, choosing a format that fits the current idea. Because voice and screen share the same lesson state, a learner can interrupt, ask for a different explanation, and continue without losing the thread.

The learning design draws on Benjamin Bloom's work on one-to-one tutoring ("The 2 Sigma Problem," Educational Researcher) and Roediger and Karpicke's research on retrieval practice ("Test-Enhanced Learning," Psychological Science, 2006). Taughtful applies those ideas through explanations that can change mid-lesson and spoken checks where learners explain ideas back.

How Taughtful.ai uses Vocal Bridge

The difficult part was not making an AI speak. Taughtful needed one agent to hold a natural conversation, stop when interrupted, and control the lesson interface at the same time. Mukul Pathak says client actions removed the team's biggest blocker:

"And then we saw something you had called client actions. And frankly, we fell in love with that. That was the single biggest blocker which we were facing with all other previous applications. The ability for the agent to do things, to control the system or your applications. So that basically cut down most of our efforts."
- Mukul Pathak, Taughtful.ai

Taughtful uses Vocal Bridge's Voice for your app surface. Vocal Bridge runs the live, interruptible conversation. Client actions let the same agent control what the Generative Player draws, so the visual changes with the spoken lesson instead of trailing behind it.

import { useEffect, useState } from "react";
import { useAgentActions } from "@vocalbridgeai/react";

export function LessonCanvas() {
  const [visual, setVisual] = useState(null);
  const { onAction } = useAgentActions();

  useEffect(() => {
    return onAction("show_slide", (payload) => setVisual(payload));
  }, [onAction]);

  return <pre>{JSON.stringify(visual, null, 2)}</pre>;
}

Illustrative. Taughtful's production player supports a richer visual vocabulary and section-aware lesson flow.

The architectural split is deliberate. Taughtful owns content ingestion, lesson planning, the Generative Player, and the learner profile. Vocal Bridge owns real-time voice, natural turn-taking, barge-in, and the Client Actions channel. Actions such as show_slide and pause_video keep the visual and spoken lesson aligned, while post-call processing gives Taughtful structured input for the learner profile.

What is hard about voice for adaptive learning?

A voice tutor cannot behave like a talking video. It has to know when the learner has taken the floor, keep a changing screen aligned with the explanation, and carry useful learning signals into the next session. Taughtful treated those as three separate system boundaries.

• Knowing when to yield is part of the teaching experience. If the tutor stops too eagerly, a normal pause breaks the lesson. If it stops too slowly, it talks over the learner at the exact moment they are confused. Taughtful reports that semantic VAD introduced enough delay to make the tutor slow to yield, so the team switched to server-side VAD and tuned vad_threshold to 0.3 and silence_duration_ms to 550. Vocal Bridge's configurable server-side VAD and barge-in handle that interruption boundary without Taughtful building its own turn-taking engine.
• The voice agent should direct the screen, not render all of it. Sending every visual through the live agent would add load and let the screen fall behind the explanation. Taughtful split the work instead. Vocal Bridge Client Actions carry lesson events such as show_slide and pause_video with the agent turn, while Taughtful's deterministic Generative Player renders the visual at the right moment.
• Adaptation has to survive the live session. Taughtful wants a later lesson to remember how someone gets unstuck, not just respond well during one call. Vocal Bridge post-call processing gives Taughtful structured session output that its learner-profile layer can reuse in the next lesson.

How do you add voice to an adaptive learning app?

Keep domain state in the application and use Vocal Bridge for the real-time conversation and control path. Taughtful's implementation suggests four practical steps:

1. Define the moments where voice should change the interface. Taughtful maps lesson events to client actions such as rendering a visual or focusing a section.
2. Treat interruption as a primary input. Barge-in is not an edge case for a tutor. It is how the learner tells the product where understanding broke.
3. Keep spoken and visual state synchronized. The action that changes the screen should come from the same agent turn that explains it.
4. Use the completed conversation after the call. Post-call processing can turn session history into preferences or application state for the next interaction.