Watch an example video here!
This takes a markdown file that you would normally convert to HTML for Substack or upload elsewhere and convert it directly into *.mp3 files for podcasts and *.mp4 files for YouTube to tap into new audiences. It uses two state-of-the-art generative AI tools/systems:
- Elevenlabs.io's new multilingual models for audio generation.
- OpenAI's DALLE3 for fun images.
It costs a few dollars a post if you publish monthly. I'm on the $22/month Elevenlabs plan based on my Substack wordcount and each post costs $2 to 8 of DALLE API credits (0.08 per widescreen image).
Expand your audience and have some fun! You can use my Elevenlabs referral link if you're going to try this out!
Notes on using generated audio: In my distribution, I make it clear that it is made via AI, you should do the same. Second, from the 11labs documentation, all audio is watermarked:
All audio generated by our model will be watermarked, so that it can be instantly traced back to us;
Second, you cannot use my generated voice per 11labs terms and how it is created. My API key in the git history is no longer valid, of course.
Right now this is not on pypi, run as following:
git clone https://github.com/natolambert/blogcaster
cd interconnects-tools
Install from requirements.txt
conda create -n media python=3.10
conda activate media
pip install -r requirements.txt
Next, you need to make sure you have subsequent accounts & API keys.
This requires an Eleven Labs account with a model id from your voice lab. This takes a few minutes to set up and is fine-tuned from 1-3minutes of recordings of yourself; very easy.
To set the OpenAI API key, add the following (from the docs) to your bashrc:
export OPENAI_API_KEY='your-api-key-here'
Similarly, for 11labs:
export ELELABS_API_KEY='your-api-key-here'
And finally for HuggingFace (optional figure storage for podcast show notes):
export HF_API_KEY='your-api-key-here'
I've begain using this directory for these images.
Additionally, the audio/visual tools I used require ffmpeg.
Blogs using these tools: (open an issue to be featured!)
| Blog | Podcast Link | YouTube Link |
|---|---|---|
| interconnects | https://podcast.interconnects.ai/ | https://www.youtube.com/@interconnects |
This is designed to work with the following data format (note, it is exactly as exported from Notion as markdown for an individual post):
scripts/
source/
└─- post-title/
| |-- post-title-name.md
| └- post-title-name/
| | img0.png
| | ...
| └─ imgN.png
| ...
└─- post-title-two/
| ...
Generate the config file (that contains the paragraphs etc)
python scripts/create-config.py source/test-post/ --date="December 24th 2023"
The image paths can be wrong if you change them on your local machine post export, double check!
Note: it is recommended to skim the config and combine things like lists, otherwise generation is split into many more parts and needlessly across images at times.
I recommend double checking file-paths for images in the config. They can be quite annoying, especially exporting from Notion.
Base usage is as follows.
python scripts/tts.py --input=source/test-post/
Audio generation returns descriptions for youtube / podcast with chapters. E.g.:
----------------------------------
Printing podcast chapter versions (does not include `see figure` audio):
----------------------------------
Interconnects year in review: 2023
The core themes of ML and the blog this year. What changes in 2024.
This is AI generated audio with Python and 11Labs. Source code can be found here: https://github.com/natolambert/interconnects-tools
Original post: https://www.interconnects.ai/p/TODO
00:00 Interconnects year in review: 2023
01:45 Brief 2024 predictions
03:37 Top posts of the year
05:07 Trends
05:11 RLHF capabilities and understanding
07:09 Open LLM ecosystem progress
08:38 LLM techniques pieces
10:12 Model releases
11:15 Moats
11:44 State of ML opinion pieces
12:33 Understanding reward/preference models
13:01 Wrap up
Non-default usage:
More paths will need to be passed to the tts.py script in the case that you're not using the same file structue and voice.
python scripts/tts.py --input=source/your-post/ --elelabs_voice='your_generative_id'
Optionally, add --farewell_audio and --figure_audio to add a farewell to every post or to tell the audience to look at the figure during the video.
Audio add music + outro (not currently using this, need to think about music more):
python experimental/add-music.py --input=audio/20231129-synthetic.mp3
python scripts/ttv-generate.py --input=source/test-post/
Optionally, one can provide a HuggingFace dataset where the images will be uploaded, so the podcast show notes can have clickable links, e.g. (--do_not_gen is for not burning extra GPUs):
python scripts/ttv-generate.py --input=examples/test-post/ --do_not_gen --hf_fig_dir=natolambert/interconnects-figures
Podcast figures:
Figure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/test-post/img_003.png
python scripts/ttv-merge.py --input=source/test-post/
Sick of workshopping 3-5minute videos no one watches? Use this tool! An example video is here and included in examples/research-talk.
This section requires imagemagick:
brew install imagemagick
Source code is provided in examples/research-talk.
Ultimately, you must download a pdf of slides from google slides / powerpoint / etc and write a script.
To create the PNG files in the right format, do the following:
convert -density 300 examples/research-talk/research-talk/talk.pdf -quality 100 examples/research-talk/images/img_%03d.png
To use the interconnects-tools:
python scripts/create-config.py examples/research-talk/
python scripts/tts.py --input=examples/research-talk/ --ignore_title
python scripts/ttv-merge.py --input=examples/research-talk/ --ignore_title
To speed up the final video to your desired length, do the math, then use the following command:
ffmpeg -i input.mp4 -filter_complex "[0:v]setpts=PTS/(SPEED_FACTOR)[v];[0:a]atempo=SPEED_FACTOR[a]" -map "[v]" -map "[a]" output.mp4
Here's an example for 10%:
ffmpeg -i examples/research-talk/video.mp4 -filter_complex "[0:v]setpts=PTS/1.1[v];[0:a]atempo=1.1[a]" -map "[v]" -map "[a]" output.mp4
Generate silence:
ffmpeg -f lavfi -i anullsrc=r=44100:cl=stereo -t 0.5 silence.mp3
- Use shorter sequeneces, 11labs tends to have volume drift on longer segments.
- Weird file paths can cause bugs still (such as () at the end of an image file).
Keeping note of features I want to add in a lazy manner:
- Adding list of figures to podcast shownotes + link because inserting them is hard.
- Verify that the try-except block in
tts.pyworks as intended. - Acronym management (so the audio is not weird).
- Additional voice for quote blocks.