We've been using WellSaid Labs for our product's onboarding tutorials and release notes for about a year now. The stock voices are excellent—clear, consistent, and they've significantly reduced our audio production time. However, our marketing team recently pushed for a more branded experience and proposed creating a custom avatar for our rather quirky brand mascot, "Terra." We went through the entire custom avatar creation process, and the result is... technically impressive but ultimately falls short for our specific use case. I wanted to detail our experience for the community, as the process involves a substantial time and financial commitment.
The creation pipeline itself is well-structured. You begin with a multi-hour recording session with a professional voice actor, following a script provided by WellSaid. This script covers a massive phonetic range. The subsequent training period took the quoted four weeks. The resulting voice is undeniably high-quality and captures the actor's timbre and pacing perfectly. Where it fails is in character. "Terra" is supposed to sound energetic, slightly mischievous, and younger. The actor we selected, based on their demo, could hit those notes in a live performance, but the trained model seems to have leveled out these peaks. The output is neutral, polished, and sonically flawless, but it lacks the specific personality we were paying to capture. It feels like we've taken a unique human voice and processed it into a superior, yet generic, text-to-speech model.
From a technical and cost perspective, here's the breakdown:
* **Initial Investment:** The custom avatar fee is substantial, squarely in the five-figure range.
* **Ongoing Costs:** You then commit to a minimum annual consumption tier for that avatar, which is several times more expensive than using the stock voices.
* **Infrastructure Integration:** We handle everything via their API. The switch from a stock voice to our custom avatar was seamless from an engineering standpoint—just a change of the `voice_id` parameter in our Terraform-provisioned Lambda function that handles audio generation.
```hcl
# Before (using stock voice)
resource "aws_lambda_function" "tts_generator" {
...
environment {
variables = {
WELLSAID_VOICE_ID = "vi_sonya"
}
}
}
# After (using custom avatar)
resource "aws_lambda_function" "tts_generator" {
...
environment {
variables = {
WELLSAID_VOICE_ID = "avatar_terra_blackwood" # Our custom avatar ID
}
}
}
```
The core issue is the return on investment. For corporate narration, explainer videos, or any content requiring authoritative clarity, this custom avatar would be a 10/10. For a brand mascot that needs character and a distinct emotional range, it's a 6/10. We're now faced with a dilemma: do we accept this high-cost, high-quality but personality-neutral voice, or revert to a carefully directed human actor for key mascot lines and use stock WellSaid for everything else? The fidelity is there, but the soul, for lack of a better term, was filtered out during training. I'm curious if others in the community have gone through the custom avatar process for animated or character-driven roles and how you've tackled this personality gap.