Files
vh c5ab99e74f kokoro: persist custom voices across container recreate
Wrapper only enumerates one voice directory (settings.voices_dir,
default /app/api/src/voices/v1_0 — inside the container's writable
layer, not bind-mounted). Override via VOICES_DIR=/app/user_voices
(host bind mount) and add a command shim that cp -r's built-ins from
the in-image v1_0 into user_voices on every start. Built-ins re-seed
fresh from the image (so upgrades that add voices propagate); custom
.pt files in user_voices are preserved (cp -r is additive).

Also adds scripts/blend_kokoro_voice.py + a playbook around it that
mirrors the wrapper's request-time voice="a(w)+b(w)" math but writes
the result as a named .pt to user_voices, making it discoverable via
GET /v1/audio/voices and persistent across recreate. Defaults to
athena = af_bella(2)+af_aoede(1) normalized.
2026-05-10 17:48:24 -07:00

94 lines
3.2 KiB
Python

#!/usr/bin/env python3
# Blend Kokoro voicepacks into a new persistent voice.
#
# Mirrors Kokoro-FastAPI's request-time blend op (see
# api/src/services/tts_service.py::_get_voices_path) so the resulting
# .pt behaves identically to the equivalent inline `voice="a(w)+b(w)"`
# request — except it lives on disk, gets a stable name, and is
# discoverable via GET /v1/audio/voices.
#
# Designed to be run INSIDE the kokoro container (where torch and the
# built-in voicepacks are already present):
#
# docker exec -u 0 kokoro python /tmp/blend_kokoro_voice.py \
# --recipe 'af_bella(2)+af_aoede(1)' --out athena
#
# I/O layout (paired with the compose changes that ship custom-voice
# persistence):
#
# * Reads from /app/user_voices/, which the compose `command:` shim
# pre-seeds with the in-image built-ins on every container start.
# So both built-ins (af_bella, af_aoede, ...) and any prior custom
# blends resolve from the same directory.
# * Writes to /app/user_voices/<out>.pt — that path is the host
# bind mount /worktank/kokoro/user_voices/, so the result is
# immediately discoverable (the wrapper has VOICES_DIR pointed
# here) AND fully persistent across container recreate, image
# upgrade, and host reboot.
import argparse
import os
import re
import sys
import torch
VOICES_DIR = "/app/user_voices"
def resolve(name: str) -> str:
p = os.path.join(VOICES_DIR, f"{name}.pt")
if os.path.exists(p):
return p
raise FileNotFoundError(f"voice {name!r} not found in {VOICES_DIR}")
def parse_recipe(recipe: str):
parts = re.split(r"([-+])", recipe.replace(" ", ""))
terms = []
for i, tok in enumerate(parts):
if i % 2 == 0:
if "(" in tok and ")" in tok:
name = tok.split("(")[0]
weight = float(tok.split("(")[1].split(")")[0])
else:
name, weight = tok, 1.0
sign = +1.0 if i == 0 or parts[i - 1] == "+" else -1.0
terms.append((name, sign * weight))
return terms
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--recipe", required=True,
help='e.g. "af_bella(2)+af_aoede(1)"')
ap.add_argument("--out", required=True, help="output voice name (no .pt)")
ap.add_argument("--no-normalize", action="store_true",
help="sum weights as-is instead of dividing by total")
ap.add_argument("--force", action="store_true",
help="overwrite existing output voice")
args = ap.parse_args()
terms = parse_recipe(args.recipe)
if not terms:
sys.exit("empty recipe")
total = sum(abs(w) for _, w in terms) if not args.no_normalize else 1.0
if total == 0:
sys.exit("weights sum to zero")
out_path = os.path.join(VOICES_DIR, f"{args.out}.pt")
if os.path.exists(out_path) and not args.force:
sys.exit(f"refusing to overwrite {out_path} (pass --force)")
blended = None
for name, w in terms:
t = torch.load(resolve(name), map_location="cpu") * (w / total)
blended = t if blended is None else blended + t
torch.save(blended, out_path)
print(f"wrote {out_path} (shape={tuple(blended.shape)}, dtype={blended.dtype})")
if __name__ == "__main__":
main()