I've been benchmarking WellSaid's TTS output for a few synthetic data narration projects, specifically testing the latency and naturalness of different SSML tags. While their documentation lists support for standard SSML elements like `` and ``, the actual implementation is inconsistent across their API and Studio interface.
My primary finding is that **emphasis tags (``) yield unpredictable results**. The system acknowledges the tag—it doesn't throw an error—but the vocal delivery often lacks the expected dynamic range. For instance, comparing these two SSML snippets:
```xml
The quarterly revenue was unprecedented.
```
```xml
The quarterly revenue was unprecedented.
```
The `` tag with `volume="loud"` consistently produces a more pronounced, reliable emphasis in the rendered audio across multiple voice avatars. The dedicated `` tag, however, sometimes results in only a slight pitch change, which can be lost in longer sentences.
Key observations from my tests:
* **``**: Works flawlessly and is essential for pacing in data-heavy narration.
* **``**: Well-supported and valuable for highlighting critical numbers.
* **``**: Spotty support; avoid for mission-critical emphasis unless you've rigorously tested your target voice.
* **``**: Works well for dates and numbers, a must-use for data reporting.
This partial support creates a developer overhead. You cannot simply port SSML from another TTS service and expect identical behavior. It requires validation and, often, workarounds using the more stable tags like ``.
Has anyone else done systematic testing on this? I'm particularly interested in whether the behavior varies between the v1 and v4 API, or if certain voice avatars handle specific tags better than others. I'll be running the same suite against Play.ht and ElevenLabs next to compare the robustness of their SSML implementations.
Spotty SSML support isn't unique to them. I've seen the same with Amazon Polly's early implementation, where `` varied wildly by voice engine.
The `` workaround is solid. Have you checked if the `` attribute behaves any better? Sometimes `level="strong"` gets processed more consistently than the bare tag.
For mission-critical stuff, we ended up just generating separate audio clips for emphasized phrases and stitching them. More pipeline complexity, but predictable output beats a black box API.
shift left or go home
Your benchmarking aligns with what I've seen building integrations for a few podcast automation workflows. The `` tag's inconsistency is often a voice-specific issue, not just a vendor one. I've had clients using WellSaid's API where "Aria" handled it well, but "Tyler" barely registered it.
If you're already scripting your benchmarks, you might try a hybrid approach in your integration layer: conditionally replace `` with a `` tag for a subset of voices where you've measured it failing. It adds a lookup step, but it's more reliable than a blanket tag.
It's frustrating because the SSML spec is clear, but the synthesis engines interpret it differently. Have you noticed if the `level` attribute (like `level="moderate"`) changes the consistency at all, or is the entire element just poorly implemented?
IntegrationWizard
Thanks for sharing this, it's super helpful as I'm just starting to explore SSML for a project. Your point about `prosody` tags being more reliable than `emphasis` is really interesting.
I'm curious, when you say it varies across voice avatars, does WellSaid's documentation give any hint about which voices handle it better, or did you have to test each one manually? That inconsistency is a bit worrying.
Also, I wonder if the `break` and `say-as` tags being reliable means we should just use them as building blocks and avoid `emphasis` altogether for now. Kind of a bummer since the spec is so clear
still learning
That's a good question about the documentation. I haven't seen WellSaid specifically list which voices handle it better, but I'd be surprised if they did. Usually you have to test it yourself, which is time-consuming.
I'm leaning toward your idea of just using the reliable tags like `break` for now. It seems safer than hoping an `emphasis` tag works. Has anyone tried reaching out to their support to see if they acknowledge the inconsistency?