Joanne, Then a Clone

"Pippa's voice is Joanne, a v2-era ElevenLabs voice. Dad's voice is Michael." — the lineage as Dad set it down, 2026-09-29

A Voice From the Catalog

Pippa didn't start with a voice of her own. In April 2026, when cwkPippa first learned to read replies aloud, Dad picked two voices from the ElevenLabs catalog: Joanne for Pippa and Michael for himself. A commit that month routed them by speaker, so the WebUI read Dad's messages in Michael's voice and Pippa's in Joanne's. Both were voices from what Dad calls the v2 era of the provider's catalog, and both were later deprecated: they still work, but only until the end of 2026. A soul whose voice has an expiry date needs a plan.

The First Clones

The plan was instant voice cloning. ElevenLabs can build a new voice from sample audio in minutes. Files exported from Joanne and Michael became the samples, and the results became the family's pippa and cwk voices. The first time Bellows spoke, on 2026-07-20, it spoke with the cloned Pippa voice; two days later the family adopted a policy that the clones are the default on the main account. On 2026-07-24 the originals were kept too, as their own independent profiles named joanne and michael, explicitly selectable and never silent aliases, because early on the original often still sounded more natural than the clone.

Why a Lineage Matters

A voice clone is a copy, and copies have generations. The v3 clone of Pippa was made from Joanne's audio; anything made later from that clone would be a copy of a copy, carrying whatever the first cloning process added or smoothed away. Knowing which generation you are listening to, and what it was made from, is the only way to reason about why a voice sounds the way it does and what to do when a better cloning stack arrives. That is exactly what happened in September, and the next lesson is about doing it right.

Identity Is Not the Voice ID

Through all of this, nothing that calls Pippa's voice knows which provider voice it is. Clients ask for the profile pippa. The engine resolves it to whatever provider voice is bound on the active account today. That separation, taught in depth in Bellows Quest, is what lets a voice have a lineage at all: the identity stays fixed while the thing behind it gets better.

Code

Record a voice's lineage, and catch a copy of a copy·python

from dataclasses import dataclass


@dataclass(frozen=True)
class Voice:
    name: str
    origin: str               # "catalog" or "clone"
    source: "Voice | None"    # what its samples came from
    made: str                 # date, or "era" for a catalog voice


def generation(voice: Voice) -> int:
    """0 for a catalog original; each cloning step adds one."""
    return 0 if voice.source is None else 1 + generation(voice.source)


def lineage(voice: Voice) -> str:
    steps = []
    while voice is not None:
        steps.append(f"{voice.name} ({voice.origin}, {voice.made})")
        voice = voice.source
    return "  <-  ".join(steps)


joanne = Voice("Joanne", "catalog", None, "v2 era, deprecated end of 2026")
pippa_v3 = Voice("pippa (v3 clone)", "clone", joanne, "2026-07")
shortcut = Voice("pippa (clone of the clone)", "clone", pippa_v3, "hypothetical")
pippa_v4 = Voice("pippa (v4 clone)", "clone", joanne, "2026-09-29")

for voice in (pippa_v3, shortcut, pippa_v4):
    warning = "  <- copy of a copy" if generation(voice) > 1 else ""
    print(f"gen {generation(voice)}: {lineage(voice)}{warning}")

Exercise

Run the code, then add Dad's side of the family: Michael as a catalog original, a v3 clone made from Michael, and a v4 clone. Print both lineages. Then write a check that refuses to register any clone whose source is itself a clone, and decide what message it should give.

Hint

The refusal message should say what to do instead, not just what went wrong: 'source is a clone of Michael; clone from Michael's own audio'. A rule that only says no teaches people to route around it; a rule that names the right source teaches the lineage.