I've been testing ElevenLabs for technical documentation narration. It consistently mangles common tech acronyms and proprietary names. "PostgreSQL" comes out as individual letters, and "Kubernetes" is a coin toss.
What's your workaround? Are you forced to spell everything out phonetically in the script, or is there a hidden pronunciation dictionary I'm missing? The marketing talks about context-aware voice, but it fails on basic B2B vocabulary.
Trust but verify.
Yeah, it's a constant battle with ElevenLabs. The "context-aware" bit seems to lean more on sentence flow than proper noun recognition. For common stuff like PostgreSQL, I've just accepted I have to write it as "Post-gres-cue-ell" in the script. Annoying, but reliable.
Have you tried their voice cloning with a sample that includes the problem terms? I got slightly better results feeding it a reference audio clip where I clearly said "Kubernetes" a few times. It's a hack, not a fix.
Honestly, for technical docs, I end up pre-processing scripts with a simple find/replace dictionary that swaps "Kubernetes" for "Koo-ber-net-ees" automatically. It feels like we're back in the early text-to-speech days!
APIs > promises
That find/replace dictionary hack is so relatable! I do the same thing with our internal project codenames at work. The real pain point for me is when you have to share scripts with a team - having "Koo-ber-net-ees" scattered throughout looks unprofessional, and you need a second, clean version for the written docs.
Have you found a tool that manages the dictionary swaps cleanly, or are you just doing it manually in a text editor? I've been using a simple Python script, but it's another step in the workflow.
The voice cloning tip is interesting. I've only used it for overall tone, not for specific term training. I'll have to try that with our most butchered product names.
Benchmark or bust
Voice cloning for specific terms never worked reliably for me. The model seems to pick up the overall voice pattern but discards the pronunciation rules from the sample.
Your find/replace dictionary is the only real solution. I pipe my markdown through a simple sed command before feeding it to the TTS engine.
sed -e 's/PostgreSQL/post-gres-cue-ell/g' -e 's/Kubernetes/koo-ber-net-ees/g' script.md
Sucks, but it's predictable. Now if only they'd let us submit a proper pronunciation lexicon.
metrics not myths
Ah, the predictable sed pipeline. It's the TTS equivalent of duct tape, isn't it? While it gets the job done, it perfectly illustrates the core absurdity we've accepted: paying a premium for "AI-driven" tech that can't parse industry-standard terms, forcing us to manually regress to phonetic spelling.
What galls me about the "submit a proper pronunciation lexicon" idea is the precedent it sets. If we, the users, do the labor of building and maintaining the lexicon for every SaaS, ERP, and IAM vendor under the sun, what exactly are we paying their machine learning teams for? We're subsidizing their data gap with our own grunt work. I've seen this play out in contract negotiations too - once you start providing the workaround, the vendor's urgency to fix the root cause evaporates. They'll just point to your functional sed script as evidence the problem is "solvable."
I wonder if the underlying issue isn't technical, but legal. Maybe they're terrified of mispronouncing a trademarked name and getting a cease-and-desist, so they default to the coward's path of letter-by-letter reading. That's the only reason I can think of for "PostgreSQL" still being butchered in 2024.
Just my 2 cents