Measurable
A closed list turns unsupported markup into a deterministic validation error.
Why the renderer changed from ElevenLabs to Fish
ElevenLabs supplied a crisp first validation contract. Fish Audio offered the more expressive destination: descriptive cues that can combine emotion, vocal posture, and physical reaction.
A measurable first contract gave way to richer descriptive performance cues.
01 / Starting point
Why ElevenLabs made sense first
The early pipeline needed a finite, machine-checkable output contract. A supplied list of vocal tags made validation straightforward: every bracketed cue could be checked against one known vocabulary before a row entered training.
Reviewable
Humans can see exactly where the teacher used a cue and correct it.
Render-ready
The target text was aimed at a real speech model rather than an abstract label taxonomy.
This was useful scaffolding. The mistake was treating the supplied list as if it were an immutable provider API enumeration rather than a project-level policy.
02 / Policy change
The restriction belonged to us, so we removed it
Several otherwise useful labels failed because they used more detailed expressive language than the closed list allowed. Once that became clear, the correct response was not to distort the labels to satisfy an invented universal rule. It was to revisit the downstream renderer.
A validator is valuable only when it protects the real product contract. If the contract changes, the validator must change with it.Renderer decision principle
The existing LoRA had already seen expressive natural-language direction. Keeping that adapter and choosing a renderer that welcomes descriptive cues preserved more of the learned signal than compressing everything to a small emotional vocabulary.
03 / Fish S2-Pro
Descriptive cues can carry the whole performance
Fish Audio S2-Pro is the planned final renderer because it can interpret meaningful natural-language tags rather than requiring every performance to collapse into a single private label. The useful unit can combine physical reaction, vocal posture, pace, pitch, and emotion.
[UGC creator opening hook, fast pace, raised pitch, slightly breathless, smiling voice, strong emphasis on the product name] I did not expect this to become the thing I reach for every afternoon.
Detailed does not mean crowded. One coherent performance instruction is usually more useful than stacking many cues that compete with each other.
What can be invented safely?
Descriptive phrases with ordinary meaning can be useful. Opaque private tag names are unreliable unless a renderer is specifically fine-tuned so that each name maps to a consistent delivery.
04 / Boundary
Renderer selection is not renderer evidence
Choosing Fish establishes a design direction, not a result. The planned same-voice A/B remains unrendered, and the public site contains no Fish credential, voice reference, or generated audio.
Renderer researchedText contract updatedComparison UI preparedFish call pendingBlind listening pending
Credential boundary
When rendering begins, provider credentials and the voice reference belong in the server-side environment only. They never enter browser storage, screenshots, or this static publication.