38 lines
1.8 KiB
TOML
38 lines
1.8 KiB
TOML
# Copy to config.toml (git-ignored) and set the paths.
|
|
|
|
[paths]
|
|
game = "C:/Program Files (x86)/Steam/steamapps/common/Enshrouded" # contains enshrouded.kfc
|
|
work = "D:/ensvoice-work" # everything derived from the game: ~15 GB at 3 seeds (takes are 24 kHz WAV)
|
|
tools = "tools" # kfc-parser.exe and emm.exe from Brabb3l/kfc-parser v0.2.0
|
|
|
|
[voice]
|
|
# reference clip of the EN narrator for the ICL clone: an id from work/refs.json (step refs);
|
|
# this one won the author's bench (MOS 3.43 vs 3.24, WER 3.7% vs 6.6%); refbench compares others
|
|
reference = "EN_Grassland_plains_Ministory_05_AlwaysTogether_Page_01_034__p1"
|
|
seeds = [1000, 1001, 1002] # takes per segment; the best one is picked automatically
|
|
extra_seeds = [1003, 1004] # only for segments with no passing take (render --weak)
|
|
|
|
[tts]
|
|
backend = "local" # "local" (qwen-tts on this GPU) or "media_node" (MCP server)
|
|
model = "siriusfreak/qwen3-tts-12hz-1.7b-ru-stress-cf" # HF id or a local folder
|
|
device = "cuda:0"
|
|
dtype = "bfloat16" # "float16" on GTX 16xx / RTX 20xx / V100 (no bf16)
|
|
batch = 8 # texts per generate call: 4 on an out-of-memory error; 16 is slower on 16 GB
|
|
|
|
# [tts.media_node] # only for backend = "media_node"
|
|
# url = "http://127.0.0.1:8082/mcp"
|
|
# voice = "enshrouded/c238" # a saved voice made from the reference clip
|
|
|
|
[qc]
|
|
mos = true # UTMOSv2 naturalness (adds ~1 s per take)
|
|
wavlm = "microsoft/wavlm-base-plus-sv"
|
|
|
|
[assemble]
|
|
# workers = 7 # QOA encoding is pure Python: CPU processes (default: cores - 1)
|
|
|
|
[mod]
|
|
id = "ru_voice_notes"
|
|
name = "RU voice: notes"
|
|
authors = []
|
|
# dir = "D:/ensvoice-work/mod-test" # build elsewhere (a test); default: <game>/mods/<id>
|