xsl
12eb1eab44
通过 Voice Design 试听选定各角色参考音色，Style Control 全量克隆并部署到 audio/mp3/tts；清理未引用音频与 audio_1 旧产物。
272 lines
8.9 KiB
Python
272 lines
8.9 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
《林夏》TTS v5 — 选定试听样本克隆 + Style Control 表演
|
|
|
|
参考音频：audio_2/tts/v4_voice_samples/{role}/*.wav（每角色保留 1 条）
|
|
输出：audio_2/tts/v5_cloned/*_wav_v5.mp3
|
|
部署：audio/mp3/tts/*_wav_v1.mp3（游戏直接可用）
|
|
|
|
用法:
|
|
VOXCPM_PORT=8002 python audio_2/scripts/04_gen_tts_clone.py --force
|
|
python audio_2/scripts/04_gen_tts_clone.py --role delivery --test
|
|
python audio_2/scripts/04_gen_tts_clone.py --no-deploy
|
|
"""
|
|
import argparse
|
|
import base64
|
|
import glob
|
|
import os
|
|
import shutil
|
|
import subprocess
|
|
import sys
|
|
import tempfile
|
|
import time
|
|
|
|
import requests
|
|
|
|
PROJECT_DIR = "/home/xsl/blind"
|
|
SAMPLE_DIR = f"{PROJECT_DIR}/audio_2/tts/v4_voice_samples"
|
|
OUT_DIR = f"{PROJECT_DIR}/audio_2/tts/v5_cloned"
|
|
DEPLOY_DIR = f"{PROJECT_DIR}/audio/mp3/tts"
|
|
HOST = os.environ.get("VOXCPM_HOST", "127.0.0.1")
|
|
PORT = os.environ.get("VOXCPM_PORT", "8002")
|
|
BASE_URL = f"http://{HOST}:{PORT}"
|
|
|
|
sys.path.insert(0, f"{PROJECT_DIR}/audio")
|
|
sys.path.insert(0, f"{PROJECT_DIR}/audio_2")
|
|
|
|
from batch_tts_voxcpm import ROLES, LINES # noqa: E402
|
|
from voice_style_prompts import get_clone_style # noqa: E402
|
|
|
|
# story.js 引用但不在 LINES 中的台词
|
|
EXTRA_LINES = [
|
|
(
|
|
"lv11_0030_self3y_voice_01",
|
|
"self_3y",
|
|
"哦对，今天是我来北京的第一百天。妈，我没跟你说，我自己数的。"
|
|
"这个城市好大。我走了很多路，见了很多人。我觉得我还没准备好，"
|
|
"但我又觉得好像准备了很久很久了。你们不用担心我。我很好。"
|
|
"以后的我，如果你听到这条，记得要好好的。",
|
|
),
|
|
(
|
|
"lv11_0030_self3y_voice_02",
|
|
"self_3y",
|
|
"我今天又没睡好。不是失眠，就是睡不踏实。"
|
|
"北京这个地方，它会让你觉得你随时都要更好，随时都要更努力。"
|
|
"我在努力。但我不知道在努力什么。"
|
|
"我就想留下一条记录。今天，今天是……算了，就今天吧。",
|
|
),
|
|
(
|
|
"lv11_0030_self3y_voice_03",
|
|
"self_3y",
|
|
"是我。23岁的我。我设了三年后打开。"
|
|
"我不知道那时候的你怎么样了。如果你很好，就当我多虑了。"
|
|
"如果你不好——你看，你熬过来了对不对。"
|
|
"你23岁的时候，也觉得很难。但你还是过来了。"
|
|
"所以……撑住啊。",
|
|
),
|
|
("lv11_sys_memo_01", "linxia", "三年前今天，一条未听的录音"),
|
|
]
|
|
|
|
os.makedirs(OUT_DIR, exist_ok=True)
|
|
os.makedirs(DEPLOY_DIR, exist_ok=True)
|
|
|
|
|
|
def find_selected_ref(role_key):
|
|
role_dir = os.path.join(SAMPLE_DIR, role_key)
|
|
wavs = sorted(glob.glob(os.path.join(role_dir, "*.wav")))
|
|
if not wavs:
|
|
return None
|
|
if len(wavs) > 1:
|
|
print(f" [警告] {role_key} 有多条参考音频，使用: {os.path.basename(wavs[0])}")
|
|
return wavs[0]
|
|
|
|
|
|
def convert_to_wav(src, dst):
|
|
subprocess.run(
|
|
["ffmpeg", "-y", "-i", src, "-ar", "16000", "-ac", "1", "-acodec", "pcm_s16le", dst],
|
|
check=True,
|
|
capture_output=True,
|
|
)
|
|
|
|
|
|
def wav_to_base64(path):
|
|
with open(path, "rb") as f:
|
|
return base64.b64encode(f.read()).decode()
|
|
|
|
|
|
def apply_speed(src, speed, dst):
|
|
if abs(speed - 1.0) < 0.01:
|
|
shutil.copy2(src, dst)
|
|
return
|
|
subprocess.run(
|
|
["ffmpeg", "-y", "-i", src, "-filter:a", f"atempo={speed}", "-acodec", "pcm_s16le", dst],
|
|
check=True,
|
|
capture_output=True,
|
|
)
|
|
|
|
|
|
def wav_to_mp3(src, dst):
|
|
subprocess.run(
|
|
[
|
|
"ffmpeg", "-y", "-i", src,
|
|
"-codec:a", "libmp3lame", "-b:a", "128k", "-ar", "24000", "-ac", "1",
|
|
dst,
|
|
],
|
|
check=True,
|
|
capture_output=True,
|
|
)
|
|
|
|
|
|
def check_server():
|
|
try:
|
|
r = requests.get(f"{BASE_URL}/health", timeout=5)
|
|
return r.status_code == 200 and r.json().get("status") == "ok"
|
|
except Exception:
|
|
return False
|
|
|
|
|
|
def register_voice(wav_path):
|
|
r = requests.post(
|
|
f"{BASE_URL}/v1/voices",
|
|
json={"wav_base64": wav_to_base64(wav_path)},
|
|
timeout=60,
|
|
)
|
|
r.raise_for_status()
|
|
return r.json()["voice_id"]
|
|
|
|
|
|
def synthesize_styled(voice_id, text, style, cfg=2.0, steps=10):
|
|
payload = {
|
|
"text": text,
|
|
"style": style,
|
|
"voice_id": voice_id,
|
|
"cfg_value": cfg,
|
|
"inference_timesteps": steps,
|
|
}
|
|
r = requests.post(f"{BASE_URL}/v1/speech/styled", json=payload, timeout=180)
|
|
r.raise_for_status()
|
|
return r.content
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(description="选定样本克隆 → v5 + 部署游戏目录")
|
|
parser.add_argument("--role", help="只合成指定角色")
|
|
parser.add_argument("--test", action="store_true", help="只合成第一条")
|
|
parser.add_argument("--dry-run", action="store_true")
|
|
parser.add_argument("--force", action="store_true")
|
|
parser.add_argument("--no-deploy", action="store_true", help="不复制到 audio/mp3/tts")
|
|
parser.add_argument("--cfg", type=float, default=2.0)
|
|
parser.add_argument("--steps", type=int, default=10)
|
|
args = parser.parse_args()
|
|
|
|
lines = LINES + EXTRA_LINES
|
|
if args.role:
|
|
lines = [(f, r, t) for f, r, t in lines if r == args.role or r.startswith(args.role)]
|
|
if args.test:
|
|
lines = lines[:1]
|
|
|
|
roles_needed = sorted(set(r for _, r, _ in lines))
|
|
|
|
print(f"\n{'=' * 60}")
|
|
print(" 《林夏》TTS v5 — 试听样本克隆")
|
|
print(f" 参考: {SAMPLE_DIR}")
|
|
print(f" 输出: {OUT_DIR}")
|
|
if not args.no_deploy:
|
|
print(f" 部署: {DEPLOY_DIR} (*_wav_v1.mp3)")
|
|
print(f" 台词: {len(lines)} 条 角色: {len(roles_needed)}")
|
|
print(f"{'=' * 60}\n")
|
|
|
|
refs = {}
|
|
for role in roles_needed:
|
|
ref = find_selected_ref(role)
|
|
if ref:
|
|
refs[role] = ref
|
|
print(f" [{role}] ← {os.path.basename(ref)}")
|
|
else:
|
|
print(f" [{role}] ✗ 未找到参考音频")
|
|
|
|
if args.dry_run:
|
|
print()
|
|
for fname, role, text in lines:
|
|
style = get_clone_style(fname, role, text) or "(无表演修饰)"
|
|
print(f" {fname} [{role}] style={style}")
|
|
print(f" ref={'yes' if role in refs else 'NO'} text={text[:40]}...")
|
|
return
|
|
|
|
missing = [r for r in roles_needed if r not in refs]
|
|
if missing:
|
|
print(f"\n[错误] 缺少参考音频: {', '.join(missing)}")
|
|
sys.exit(1)
|
|
|
|
if not check_server():
|
|
print(f"[错误] VoxCPM2 未就绪: {BASE_URL}/health")
|
|
sys.exit(1)
|
|
|
|
tmp = tempfile.mkdtemp(prefix="voxcpm5_")
|
|
voice_ids = {}
|
|
for role, ref_path in sorted(refs.items()):
|
|
wav = os.path.join(tmp, f"{role}_ref.wav")
|
|
convert_to_wav(ref_path, wav)
|
|
print(f" 注册音色: {role} ... ", end="", flush=True)
|
|
voice_ids[role] = register_voice(wav)
|
|
print(f"ok ({voice_ids[role][:10]}...)")
|
|
|
|
ok = skip = fail = 0
|
|
manifest_path = os.path.join(OUT_DIR, "clone_manifest.tsv")
|
|
manifest_lines = ["filename\trole\tref\tstyle\ttext\n"]
|
|
|
|
for i, (fname, role, text) in enumerate(lines, 1):
|
|
out_v5 = os.path.join(OUT_DIR, f"{fname}_wav_v5.mp3")
|
|
out_game = os.path.join(DEPLOY_DIR, f"{fname}_wav_v1.mp3")
|
|
style = get_clone_style(fname, role, text)
|
|
ref_name = os.path.basename(refs[role])
|
|
manifest_lines.append(
|
|
f"{fname}\t{role}\t{ref_name}\t{(style or '自然说话').replace(chr(9), ' ')}\t"
|
|
f"{text.replace(chr(9), ' ')}\n"
|
|
)
|
|
|
|
if os.path.exists(out_v5) and not args.force:
|
|
print(f" [{i:3d}/{len(lines)}] SKIP {fname}")
|
|
skip += 1
|
|
continue
|
|
|
|
if role not in voice_ids:
|
|
print(f" [{i:3d}/{len(lines)}] MISS {fname}")
|
|
fail += 1
|
|
continue
|
|
|
|
print(f" [{i:3d}/{len(lines)}] {fname} [{role}] ... ", end="", flush=True)
|
|
t0 = time.time()
|
|
try:
|
|
wav_bytes = synthesize_styled(
|
|
voice_ids[role], text, style, cfg=args.cfg, steps=args.steps
|
|
)
|
|
raw_wav = os.path.join(tmp, f"{fname}_raw.wav")
|
|
spd_wav = os.path.join(tmp, f"{fname}_spd.wav")
|
|
with open(raw_wav, "wb") as f:
|
|
f.write(wav_bytes)
|
|
speed = ROLES.get(role, {}).get("speed", 1.0)
|
|
apply_speed(raw_wav, speed, spd_wav)
|
|
wav_to_mp3(spd_wav, out_v5)
|
|
if not args.no_deploy:
|
|
shutil.copy2(out_v5, out_game)
|
|
kb = os.path.getsize(out_v5) // 1024
|
|
print(f"ok ({time.time() - t0:.1f}s, {kb}KB)")
|
|
ok += 1
|
|
except Exception as e:
|
|
print(f"FAIL: {e}")
|
|
fail += 1
|
|
|
|
with open(manifest_path, "w", encoding="utf-8") as f:
|
|
f.writelines(manifest_lines)
|
|
|
|
print(f"\n完成: {ok} ok, {skip} skip, {fail} fail")
|
|
print(f"输出: {OUT_DIR}")
|
|
print(f"清单: {manifest_path}")
|
|
if not args.no_deploy:
|
|
print(f"游戏目录: {DEPLOY_DIR}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|