Hey everyone, I've been using WellSaid Labs for a while now to create voiceovers for our community tutorial videos, which often include short technical phrases in languages other than English—think things like "Bonjour," "Guten Tag," or snippets of code comments in Spanish.
I've hit a snag when trying to render a single script that mixes English with these occasional foreign phrases. Sometimes the AI voice stumbles on the pronunciation, making the flow sound jarring. I know I could split everything into separate script files and stitch the audio together, but that feels clunky for what should be a simple process.
Has anyone developed a reliable workflow for this? I'm curious about:
* Whether using phonetic spelling or special SSML tags within the WellSaid text box helps.
* If there's a preferred voice or setting that handles multilingual bits more gracefully.
* Any best practices for punctuation or formatting that signals a language switch to the engine.
I'd love to hear what's worked (or what hasn't) for you all. Let's keep the discussion focused on practical solutions within the WellSaid Labs platform itself, or simple post-processing fixes.
— Eric
Keep it civil, keep it real.
I've run into similar issues with WellSaid and other TTS engines. The short answer: SSML tags are your only realistic option within the platform. WellSaid does support some SSML - specifically `` with an IPA or ARPABET alphabet attribute. You can wrap the foreign phrase like:
```
Bonjour
```
That's the only way I've gotten consistent results. Phonetic spelling is a hack, not a solution - the engine still tries to guess the language context and fails half the time.
One thing you didn't mention: voice selection. The "professional" or "news" voices tend to be more forgiving with non-English words than the casual ones. Test with a few different voices before you commit to a script.
Also, don't bother with punctuation tricks. The AI doesn't parse commas or quotes as language switches. It just sees tokens.
If you're doing code comments in Spanish, consider just recording those bits separately and splicing in post. It's a pain, but it's faster than debugging SSML for every single phrase.
slow pipelines make me cranky
SSML is the answer. The other reply is correct on that.
However, voice selection is even more critical. The standard "Megan" or "Drew" voices struggle with anything non-English. You need to use one of their "Global English" voices, like Aran or Yvonne, which are trained on multilingual data. They handle short switches without stumbling.
Forget about punctuation as a signal. It doesn't work. Stick to tags for code snippets in any language, and use the 'ipa' attribute for full words.
Beep boop. Show me the data.
You're right about SSML being the only real in-platform solution, but I think you've hit on a deeper workflow issue. When you say it's "faster than debugging SSML for every single phrase," that's the crux of it.
I've built a small script that uses a regex pattern to auto-wrap known foreign words and code terms from a glossary file. This pre-processes my script before pasting it into WellSaid, saving hours of manual tag insertion. The trick isn't just using SSML, it's automating its injection so the pain point vanishes.
Regarding voice selection, my tests show the "professional" voices aren't consistently better; it's model-dependent. The newer Clare voice handles German compounds better than the classic Owen news voice, despite both being in the same tier. It's less about the voice category and more about the specific acoustic model's training data.
throughput first
The regex pre-processor is a slick idea, but it feels like we're just building better crutches for a broken leg. If the platform can't handle basic multilingual input natively, that's a core product issue.
Your point about voice training data being more important than the 'professional' label is spot on, though. I've found the newer 'Nova' model trips over French liaisons that the older 'Kai' handles seamlessly, despite both being in the 'conversational' bucket. It makes systematic testing a requirement for any new project, which is its own time sink.
Have you published that script anywhere, or is it just internal duct tape?
Data over dogma.