Feature: Self-hosted multi-input transcript/translation agent pipeline

[ydangishere]

Duplicates

• I have searched the existing issues

Summary 💡

Problem

There is currently no reusable pipeline for handling transcript workflows such as:

• browser media translation
• webinar/live translation
• meeting transcription + note-taking

Proposal

Introduce a self-hosted / bring-your-own-key agent pipeline.

AutoGPT provides:

• block/UI
• orchestration
• integration

Users provide:

• runtime (local/VPS)
• API keys (STT/LLM)
• infrastructure

Input Modes

1. Transcript (YouTube)

• use existing captions
• no STT
• lower cost

2. Audio Stream

• mic / browser
• STT required

Pipeline

Input → Text → Translate/Summarize → Output

MVP

Phase 1:

• YouTube transcript input

Phase 2:

• Audio input

Notes

This is not a platform-hosted realtime audio service.

Examples 🌈

Example use cases:

1. YouTube video translation
   User provides a YouTube URL → system fetches transcript → translates into target language → outputs readable text or subtitles.

2. Browser media translator
   User captures audio from a browser tab → converts speech to text → translates in near real-time → displays live text.

3. Meeting assistant
   User records meeting audio → transcribes speech → summarizes key points → outputs structured notes.

4. Webinar/live stream translation
   Audio stream → STT → translation → live subtitle-style output.

Motivation 🔦

Currently, there are separate blocks and tools for speech-to-text, translation, and text processing, but no unified pipeline that connects them into a reusable workflow.

This makes it difficult to build real-world use cases such as:

• live translation
• meeting transcription + note-taking
• media content translation

A multi-input pipeline (audio + transcript) would simplify these workflows and allow users to build practical AI agents without manually wiring multiple components.

This also enables cost optimization by allowing users to use existing transcripts when available instead of always relying on STT.

[ydangishere]

Thanks for sharing.

This proposal is currently focused on building a transcript/translation pipeline (audio/transcript → text → processing → output).

Happy to explore integrations later if relevant, but for now I’d like to keep the scope focused on the core pipeline.

[ydangishere]

Hi, I’ve built a minimal standalone MVP to validate the transcript input direction before moving to full pipeline integration.

Repo branch:
https://github.com/ydangishere/AutoGPT/tree/feature/transcript-mvp-minimal/autogpt/experiments/transcript_mvp_minimal

What this MVP does:

• Input: YouTube URL
• Output: transcript text
• Uses Supadata API for reliability (direct YouTube transcript access can fail due to IP/rate limits)
• No API key is committed (uses SUPADATA_API_KEY env variable)

Run locally:

1. Get API key from https://supadata.ai
2. Set environment variable

Windows:
set SUPADATA_API_KEY=your_key

Mac/Linux:
export SUPADATA_API_KEY=your_key

3. Run:
   python autogpt/experiments/transcript_mvp_minimal/main.py

Goal:
Validate transcript ingestion as a reusable input step for the pipeline (audio/transcript → text → processing → output)

Would appreciate feedback on whether this direction makes sense before proceeding to full block integration (and later audio support).