The WeaveSignal contributors
16833c96ae
- tests/demo-video/: the script, designed voices, the run through the real desk and the edit that make the website's demo video (the media themselves are not in git). - The home page plays the video, with its credits. - deploy/cloudron/CHANGELOG.md: one line per entry; CloudronVersions.json rebuilt.
72 lines
3 KiB
Mojo
72 lines
3 KiB
Mojo
# SPDX-License-Identifier: AGPL-3.0-only
|
|
# SPDX-FileCopyrightText: 2026 The WeaveSignal contributors
|
|
#
|
|
# Designs one voice per character of the demo video (tests/demo-video/script.json): OmniVoice's
|
|
# voice design mode, which makes a voice from a description (gender, age, pitch, accent) and a
|
|
# seed, with no reference recording and so no real person. Each character's design text is read
|
|
# in its designed voice and written as <out>/<id>.wav with its transcript in <out>/<id>.txt;
|
|
# synth.sh then reads every line through VoiceLab's tts-serve with that clip as the reference
|
|
# and the same seed, so a character keeps one voice across all their lines (tts-serve's
|
|
# OmniVoice server has no voice-design field of its own).
|
|
#
|
|
# Run through design.sh, which points Mojo at the VoiceLab OmniVoice environment. Only one
|
|
# OmniVoice may hold the GPU at a time: run this with voicelab-tts stopped.
|
|
#
|
|
# tests/demo-video/design.sh [seed-offset]
|
|
from std.python import Python, PythonObject
|
|
from std.sys import argv
|
|
|
|
|
|
def main() raises:
|
|
var args = argv()
|
|
if len(args) < 3:
|
|
raise Error("usage: design.mojo <script.json> <out-dir> [seed-offset]")
|
|
var script_path = String(args[1])
|
|
var out_dir = String(args[2])
|
|
var offset = 0
|
|
if len(args) > 3:
|
|
offset = Int(String(args[3]))
|
|
|
|
# Mojo's own launcher is a user-site Python program: keep that site's packages out of
|
|
# this interpreter, so only the OmniVoice environment's are imported.
|
|
var sys = Python.import_module("sys")
|
|
var kept = Python.list()
|
|
for entry in sys.path:
|
|
if not String(entry).__contains__("/.local/lib/python"):
|
|
kept.append(entry)
|
|
sys.path = kept
|
|
var json = Python.import_module("json")
|
|
var os = Python.import_module("os")
|
|
var random = Python.import_module("random")
|
|
var np = Python.import_module("numpy")
|
|
var torch = Python.import_module("torch")
|
|
var sf = Python.import_module("soundfile")
|
|
var builtins = Python.import_module("builtins")
|
|
var omnivoice = Python.import_module("omnivoice")
|
|
|
|
var script = json.load(builtins.open(script_path))
|
|
var cast = script["cast"]
|
|
var text = script["design_text"]
|
|
os.makedirs(out_dir, exist_ok=True)
|
|
|
|
print("loading OmniVoice")
|
|
var model = omnivoice.OmniVoice.from_pretrained(
|
|
"k2-fsa/OmniVoice", device_map="cuda", torch_dtype=torch.bfloat16
|
|
)
|
|
var rate = model.sampling_rate
|
|
for key in builtins.list(cast.keys()):
|
|
var who = cast[key]
|
|
var seed = Int(py=who["seed"]) + offset
|
|
random.seed(seed)
|
|
np.random.seed(seed)
|
|
torch.manual_seed(seed)
|
|
torch.cuda.manual_seed_all(seed)
|
|
var audios = model.generate(
|
|
text=text, language="en", instruct=who["voice"]
|
|
)
|
|
var wav = String(out_dir, "/", String(key), ".wav")
|
|
sf.write(wav, audios[0], rate)
|
|
var f = builtins.open(String(out_dir, "/", String(key), ".txt"), "w")
|
|
_ = f.write(text)
|
|
_ = f.close()
|
|
print(String(key), seed, String(who["voice"]), "->", wav)
|