Automate YouTube video creation with GPT-4 and ElevenLabs

To automate YouTube video creation using ai video creation workflows, combine GPT-4 for script generation, ElevenLabs for voice synthesis, and FFmpeg for video assembly. GPT-4 writes scripts from topic keywords; ElevenLabs converts those scripts to professional audio tracks with multiple voice options; FFmpeg stitches audio with stock footage or generated visuals into a final MP4. You orchestrate this pipeline in n8n or Make, triggering on a schedule or webhook. The result: fully produced videos ready to upload to YouTube without manual editing. This approach scales from one video per week to dozens daily, cutting production time from hours to minutes per video while maintaining consistent quality and branding across your channel.

What you need

To build an automated ai video creation pipeline, you'll need a text-to-speech engine, a video composition tool, storage, and a publishing integration. Here's what works:

How it works

1. Script generation with GPT-4. You send a topic or outline to GPT-4 via API (or use n8n/Make to trigger it on a schedule). GPT-4 returns a structured script with scene descriptions, dialogue, and timing cues - typically 500-1500 words for a 5-10 minute video.

1. Voice synthesis via ElevenLabs. Pass the script text to ElevenLabs text-to-speech API (from $0.30 per 1k characters for standard voices; check current pricing for premium tiers). ElevenLabs returns a WAV or MP3 file with natural prosody and optional voice cloning, which you store in Google Cloud Storage or AWS S3.

1. Scene asset generation. Use the scene descriptions from GPT-4 to query an image API (DALL-E, Midjourney API, or Stability AI) or pull stock footage from Pexels/Pixabay. Store frames or clips in your cloud bucket alongside the audio.

1. Video assembly with FFmpeg. Use FFmpeg to concatenate images, video clips, and the ElevenLabs audio track into a single MP4 file. FFmpeg runs locally or in a containerized worker (AWS Lambda or Google Cloud Run) and outputs H.264 video at 1080p or 4K.

1. Upload to YouTube. Use the YouTube Data API v3 to publish the finished MP4 directly to your channel with auto-generated title, description (also from GPT-4), and tags. Schedule publishing or go live immediately.

How to build it

This workflow chains GPT-4 for script generation, ElevenLabs for voice synthesis, and ffmpeg for video assembly. The orchestration runs in n8n, with video files stored on Google Cloud Storage and optional AWS Lambda for parallel processing.

1. Set up n8n and create a new workflow. Start with a manual trigger node or a webhook to accept a video topic (e.g., "AI trends in 2024"). Add an HTTP Request node to call the GPT-4 API. In the node, set Method to POST, URL to https://api.openai.com/v1/chat/completions, and add an Authorization header with your OpenAI API key. In the Body, pass a system prompt that instructs GPT-4 to generate a short YouTube script (300-500 words, conversational tone, with clear section breaks).

1. Extract the script from the GPT-4 response. Use a Set node to parse the response JSON and isolate the script text. Map the output to a variable like {{ $json.choices[0].message.content }}. This becomes your source text for voice synthesis.

1. Call ElevenLabs to generate audio. Add another HTTP Request node. Set Method to POST and URL to https://api.elevenlabs.io/v1/text-to-speech/{voice_id}, where {voice_id} is one of ElevenLabs' preset voices (e.g., 21m00Tcm4TlvDq8ikWAM for a male voice). Add the header xi-api-key: YOUR_ELEVENLABS_API_KEY. In the Body, pass { "text": "{{ $json.script }}", "model_id": "eleven_monolingual_v1" }. ElevenLabs returns a binary audio file; save it to a temporary variable or directly to Google Cloud Storage.

1. Upload the audio file to Google Cloud Storage. Use the Google Cloud Storage node in n8n (or an HTTP Request with signed URLs). Specify your bucket name and file path (e.g., videos/audio_{timestamp}.mp3). This centralizes storage and makes the file accessible for the next step.

1. Generate or source a background video. For simplicity, either use a stock video library API (Pexels, Pixabay) or pre-upload a generic background video to Google Cloud Storage. If you want dynamic visuals, call an image generation API (DALL-E, Midjourney API) to create frames, but this adds latency. For MVP, a static background works.

6. Assemble video with ffmpeg using a Lambda function or local execution. Create a small AWS Lambda function (or run ffmpeg on a server) that: - Downloads the audio from Google Cloud Storage - Downloads the background video - Overlays text captions from the script - Merges audio and video into an MP4

The ffmpeg command looks like:

 ffmpeg -i background.mp4 -i audio.mp3 \
 -filter_complex "[0:v]scale=1920:1080[v];[v][1:a]concat=n=1:v=1:a=1[out]" \
 -map "[out]" -c:v libx264 -c:a aac -shortest output.mp4
 

Wrap this in a Lambda handler that accepts S3 URIs as input and outputs the final MP4 to a public S3 bucket or back to Google Cloud Storage.

1. Invoke the Lambda function from n8n. Add an AWS Lambda node (or HTTP Request to invoke the Lambda endpoint). Pass the audio URI and background video URI as parameters. Wait for the function to complete and return the output video URL.

1. Upload the final video to YouTube. Use the YouTube Data API v3. Add an HTTP Request node with Method POST to https://www.googleapis.com/youtube/v3/videos?part=snippet,status. Include your OAuth 2.0 access token in the Authorization header. In the Body, set the video title, description, tags, and privacy status. Then use multipart/form-data to attach the MP4 file from Google Cloud Storage or Lambda output.

1. Log the YouTube URL and clean up. Extract the video ID from the YouTube API response and store it in a database or send it to Slack. Delete temporary files from Google Cloud Storage to avoid storage bloat.

Here is a minimal n8n workflow excerpt (JSON) showing steps 1-3:

```json { "nodes": [ { "parameters": { "method": "POST", "url": "https://api.openai.com/v1/chat/completions", "authentication": "predefinedCredentialType", "nodeCredentialType": "openaiApi", "jsonParameters": true, "body": { "model": "gpt-4", "messages": [ { "role": "system", "content": "You are a YouTube scriptwriter. Generate a 400-word script for a video about {{ $json.topic }}. Use a conversational tone. Include an intro, 3 main points, and a call-to-action." }, { "role": "user", "content": "Topic: {{ $json.topic }}" } ], "temperature": 0.7, "max_tokens": 1000 } }, "name": "GPT-4 Script", "type": "n8n-nodes-base.httpRequest", "typeVersion": 4.1, "position": [250, 300] }, { "parameters": { "method": "POST", "url": "https://api.elevenlabs.io/v1/text-to-speech/21m00Tcm4TlvDq8ikWAM", "headers": { "xi-api-key": "{{ $env.ELEVENLABS_API_KEY }}" }, "jsonParameters": true, "body": { "text": "{{ $json.script }}", "model_id": "eleven_monolingual_v1" } }, "name": "

What it costs to run

Assumptions: GPT-4 API at $0.03/1K input, $0.06/1K output tokens; ElevenLabs at $0.30/1K characters (Standard tier, check current pricing); FFmpeg runs on local hardware or single t3.micro EC2 instance; videos average 3-5 minutes; no concurrent processing overhead; YouTube hosting is free.

Where this breaks

Voice generation latency exceeds YouTube upload window. ElevenLabs text-to-speech can take 10-30 seconds per minute of audio, especially on free tier or during peak hours. If your workflow queues 50+ videos, the entire batch stalls. Pre-generate all voice audio in a separate n8n or Make workflow that runs 24 hours before the video assembly step, or switch to a cached voice model tier on ElevenLabs (check current pricing for priority queuing).

GPT-4 generates scripts with timing mismatches. The model outputs 2000-word scripts without considering pacing; ElevenLabs then produces 8-12 minutes of audio while your ffmpeg timeline expects 4 minutes. The video has dead air or compressed speech. Prompt GPT-4 with a hard constraint: "Write exactly 450 words at 130 words per minute for a 3.5-minute video" and validate output word count before passing to ElevenLabs.

YouTube API quota exhaustion mid-batch. Google Cloud's YouTube Data API grants 10,000 quota units per day; a single video upload (with title, description, tags, thumbnail) costs 1,500 units. A 10-video batch hits the ceiling by video 7. Spread uploads across multiple service accounts or use YouTube's resumable upload protocol via ffmpeg and the Google Cloud SDK to retry failed chunks without re-consuming quota.

ffmpeg encoding fails silently on AWS Lambda timeout. Video transcoding to H.264 for YouTube compliance takes 2-5 minutes per 10-minute source; AWS Lambda's 15-minute timeout is tight, and network I/O to S3 can trigger premature termination. Use AWS Elemental MediaConvert instead (a managed transcoding service) or run ffmpeg on an EC2 instance with a 30-minute timeout and SQS job queuing to decouple encoding from upload logic.

Can I use GPT-4 to write scripts for YouTube videos automatically?

Yes. GPT-4 can generate full video scripts from a prompt specifying topic, tone, length, and target audience. Feed the output directly into ElevenLabs for voice generation, then use ffmpeg to sync audio with stock footage or slides - this is the core of most automated ai video creation workflows.

How do I handle lip-sync if I'm using AI-generated voiceovers?

You don't need to. Most automated YouTube workflows use static visuals (slides, B-roll, text overlays, or animated graphics) paired with ElevenLabs audio rather than talking-head video. If you need lip-sync, you'll need a separate tool like Synthesia or D-ID, which adds cost and latency; static visuals are faster and cheaper.

What's the cheapest way to host and encode videos before uploading to YouTube?

AWS S3 + ffmpeg on an EC2 instance (or Google Cloud Storage + Cloud Run) costs roughly $0.023/GB for storage and $0.015 per minute of transcoding. For low volumes (under 10 videos/month), running ffmpeg locally on your machine is free; for higher volumes, Google Cloud's Transcoder API or AWS MediaConvert become cost-effective at scale.

Do I need a Google Cloud or AWS account just to upload to YouTube?

No. YouTube's API and upload flow work independently of Google Cloud or AWS. You only need those platforms if you're automating encoding, storage, or thumbnail generation at scale; for simple script-to-voice-to-video workflows, you can encode locally with ffmpeg and upload directly via the YouTube Data API using a service account.

For a deeper technical reference, see n8n's documentation.

Related reading

• Launch a faceless AI content automation bot that scrapes niche news, builds explainer videos with ElevenLabs, and publishes to YouTube Shorts automatically
• voice ai for appointment booking: Build a Vapi + ElevenLabs Agent that books slots and sends reminders 24/7
• How to make faceless AI videos for TikTok using Pika Labs and ElevenLabs