My team won't read a 50-page PDF. So I pipe them into a daily "competitor intel" podcast. Speechify is the TTS engine in the pipeline.
Here's the workflow:
1. Scrape/Clean new whitepapers to markdown.
2. Use a python script to chunk them and prepend a standard intro/outro.
3. Feed the text to Speechify's API, using the `premium` voice. (The default ones sound like a bored robot, which, fair, but not for sales).
4. ArgoCD syncs the generated MP3s to an S3 bucket. The sales team's app polls the bucket.
The key is the API integration. Basic config for the batch job looks like this:
```yaml
# In the K8s CronJob spec
env:
- name: SPEECHIFY_API_KEY
valueFrom:
secretKeyRef:
name: speechify-secrets
key: apikey
- name: VOICE_ID
value: "premium_male_01"
```
It's not perfect. Complex tables or diagrams are lost, obviously. But for dense technical prose, it works. The sales guys get the gist while commuting. Better than the docs collecting digital dust.
This is brilliant, especially the part about the sales team getting the gist while commuting. A daily audio update sounds way more digestible.
How do you handle the chunking in your python script? I've tried something similar for email summaries, and making the breaks sound natural is tricky.
Do you ever add a quick summary or key takeaway at the start? Might help if they only have a few minutes.
That's a really good point about the summary at the start. I hadn't considered that, but it makes sense for someone who only has time for the high points. In my own work with ERP documentation, I've found that even a one-sentence context can drastically change how someone interprets the following details.
On your question about chunking for natural breaks, I'm also curious. My instinct would be to split at major section headers or after a complete thought paragraph, but I've seen TTS engines sometimes create awkward pauses even there. Do you insert a brief silence between chunks, or does the script try to stitch the audio files back together seamlessly?
Also, how do you handle references to charts or figures in the whitepaper? That's always the tricky part when converting a visual document into audio.
That's a clever approach to get around the "TL;DR" problem. I'm currently trying to get my own team to adopt something similar for migrating old system docs into training materials.
Question on the scraping step: how reliable is your PDF-to-markdown process? I've found legacy documents often have weird formatting that breaks the structure, and then the audio ends up with jumbled sentences. Did you have to write a lot of custom cleanup rules, or are you using a particular library that handles it well?
One step at a time