eric/gitlab-instance-0a899031_douyinyoutube
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

's'

Browse Source
This commit is contained in:
eric
2026-03-19 03:21:09 -05:00
parent 1a2b43fd11
commit 7a03a670bc
3 changed files with 30 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
config/youtube.ini
View File

@@ -18,6 +18,9 @@ enable_demucs = 是
# 是 = 有显卡则用 GPU；否 = 强制 CPU；无显卡时自动回退到 CPU
demucs_use_gpu = 是
# 人声混合比例（0.5-1.0），改善人声自然度，减少尖锐感
# 0.85 = 85%人声+15%其他（推荐，更自然）；1.0 = 纯人声（可能尖锐）
demucs_vocal_mix = 0.85
# 人声混合比例（0.5-1.0），改善自然度、减少尖锐
# 0.75 = 75%人声+25%混合（推荐）；1.0 = 纯人声
demucs_vocal_mix = 0.75
# 混合来源：bass=低音（无混响无重复）；other=其他（可能带混响导致重复感）
demucs_mix_source = bass

31
demucs_pipeline.py
View File

@@ -75,10 +75,11 @@ DEMUCS_CHUNK_SEC = 2.0
BYTES_PER_SAMPLE = 2
def _demucs_separate_vocals(model, audio_chunk, device, vocal_mix_ratio=1.0):
def _demucs_separate_vocals(model, audio_chunk, device, vocal_mix_ratio=1.0, mix_source="bass"):
"""
对音频块进行人声分离，返回人声轨道。
vocal_mix_ratio: 1.0=纯人声；0.85=85%人声+15%other（更自然，减少尖锐感）
vocal_mix_ratio: 1.0=纯人声；0.75=75%人声+25%混合（更自然）
mix_source: "bass"=与低音混合(少混响无重复感)；"other"=与其他混合(可能带混响)
"""
import torch
# audio_chunk: numpy or tensor, shape (channels, samples)
@@ -93,17 +94,17 @@ def _demucs_separate_vocals(model, audio_chunk, device, vocal_mix_ratio=1.0):
with torch.no_grad():
out = apply_model(model, wav, device=device, progress=False, overlap=0.25)
# out: (batch, sources, channels, samples)
# sources: [drums, bass, other, vocals]
# sources: [drums=0, bass=1, other=2, vocals=3]
vocals = out[0, 3]
if vocal_mix_ratio >= 1.0:
return vocals.cpu().numpy()
# 混合 other 以保留更自然的人声，减少纯人声的尖锐感
other = out[0, 2]
mixed = vocals * vocal_mix_ratio + other * (1.0 - vocal_mix_ratio)
# 混合 bass 或 other：bass 无混响、无重复感；other 可能带混响导致"重复一遍"
blend = out[0, 1] if mix_source == "bass" else out[0, 2]
mixed = vocals * vocal_mix_ratio + blend * (1.0 - vocal_mix_ratio)
return mixed.cpu().numpy()
def _run_demucs_worker(audio_read_pipe, audio_write_socket, model, device, stop_event, vocal_mix_ratio=1.0):
def _run_demucs_worker(audio_read_pipe, audio_write_socket, model, device, stop_event, vocal_mix_ratio=1.0, mix_source="bass"):
"""Demucs 工作线程：从 FFmpeg 读音频，分离人声，写入 socket"""
import numpy as np
chunk_samples = int(DEMUCS_SAMPLE_RATE * DEMUCS_CHUNK_SEC * DEMUCS_CHANNELS)
@@ -131,7 +132,7 @@ def _run_demucs_worker(audio_read_pipe, audio_write_socket, model, device, stop_
arr = np.frombuffer(chunk, dtype=np.int16)
arr = arr.reshape(-1, DEMUCS_CHANNELS).T.astype(np.float32) / 32768.0
try:
vocals = _demucs_separate_vocals(model, arr, device, vocal_mix_ratio)
vocals = _demucs_separate_vocals(model, arr, device, vocal_mix_ratio, mix_source)
# float -> s16le
out = (vocals * 32768).astype(np.int16).tobytes()
try:
@@ -150,7 +151,7 @@ def _run_demucs_worker(audio_read_pipe, audio_write_socket, model, device, stop_
try:
arr = np.frombuffer(buf, dtype=np.int16)
arr = arr.reshape(-1, DEMUCS_CHANNELS).T.astype(np.float32) / 32768.0
vocals = _demucs_separate_vocals(model, arr, device, vocal_mix_ratio)
vocals = _demucs_separate_vocals(model, arr, device, vocal_mix_ratio, mix_source)
out = (vocals * 32768).astype(np.int16).tobytes()
audio_write_socket.sendall(out)
except Exception:
@@ -178,7 +179,8 @@ def run_demucs_pipeline(
stream_ended_exception,
cleanup_proc_fn,
stderr_reader_fn,
vocal_mix_ratio: float = 0.85,
vocal_mix_ratio: float = 0.75,
mix_source: str = "bass",
stats_interval_sec: int = 10,
):
"""
@@ -194,8 +196,9 @@ def run_demucs_pipeline(
device = next(model.parameters()).device
device_str = "cuda" if "cuda" in str(device) else "cpu"
mix_info = "纯人声" if vocal_mix_ratio >= 1.0 else f"{int(vocal_mix_ratio*100)}%人声+{int((1-vocal_mix_ratio)*100)}%其他"
print(f"[INFO] Demucs 人声分离已启用 ({device_str})，模式: {mix_info}")
src_name = "低音" if mix_source == "bass" else "其他"
mix_info = "纯人声" if vocal_mix_ratio >= 1.0 else f"{int(vocal_mix_ratio*100)}%人声+{int((1-vocal_mix_ratio)*100)}%{src_name}"
print(f"[INFO] Demucs 人声分离已启用 ({device_str})，模式: {mix_info}，低通 9kHz 降尖锐")
# TCP 服务器用于音频（跨平台）
server = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
@@ -271,7 +274,7 @@ def run_demucs_pipeline(
"-x264-params", "force-cfr=1:scenecut=0:open_gop=0:sync-lookahead=0",
"-pix_fmt", "yuv420p",
"-c:a", "aac", "-b:a", "128k", "-ar", "48000", "-ac", "2",
"-af", "aresample=async=1:first_pts=0",
"-af", "lowpass=f=9000,aresample=async=1:first_pts=0",
"-flags", "+global_header", "-flvflags", "no_duration_filesize",
"-f", "flv", youtube_rtmp,
]
@@ -339,7 +342,7 @@ def run_demucs_pipeline(
# 注意：proc_a 输出到 stdout，我们读。conn 是 FFmpeg 作为 client 连上来的，所以 FFmpeg 在读，我们写 conn。
demucs_thread = threading.Thread(
target=_run_demucs_worker,
args=(proc_a.stdout, conn[0], model, device, stop_event, vocal_mix_ratio),
args=(proc_a.stdout, conn[0], model, device, stop_event, vocal_mix_ratio, mix_source),
daemon=True,
)
demucs_thread.start()

9
douyin_youtube_ffplay.py
View File

@@ -104,7 +104,8 @@ def start_douyin_youtube_ffplay(
enable_demucs = False
demucs_use_gpu = True
demucs_vocal_mix = 0.85 # 0.85=人声更自然，1.0=纯人声（可能尖锐）
demucs_vocal_mix = 0.75 # 0.75=更自然少尖锐；1.0=纯人声
demucs_mix_source = "bass" # bass=无混响无重复；other=可能带混响
video_bitrate_mbps = 4.5 # 优先画质，YouTube 720p 推荐 4-4.5
stats_interval_sec = 10 # 推流状态日志间隔
if os.path.exists(config_file):
@@ -125,6 +126,9 @@ def start_douyin_youtube_ffplay(
demucs_vocal_mix = v
except ValueError:
pass
elif key_lower == "demucs_mix_source":
if val.lower() in ("bass", "other"):
demucs_mix_source = val.lower()
elif key_lower == "video_bitrate_mbps":
try:
v = float(val)
@@ -145,7 +149,7 @@ def start_douyin_youtube_ffplay(
# 兼容：若上面未匹配到 key，取第一个非配置项的值作为 stream key
if youtube_rtmp == rtmp_base + default_key:
for k, v in config.items("youtube"):
if k.strip().lower() not in ("enable_demucs", "demucs_use_gpu", "demucs_vocal_mix", "video_bitrate_mbps", "stats_interval_sec") and v.strip():
if k.strip().lower() not in ("enable_demucs", "demucs_use_gpu", "demucs_vocal_mix", "demucs_mix_source", "video_bitrate_mbps", "stats_interval_sec") and v.strip():
v = v.strip()
youtube_rtmp = v if v.startswith("rtmp://") else rtmp_base + v
print(f"[INFO] 从 {config_file} 读取 YouTube stream key: {youtube_rtmp}")
@@ -532,6 +536,7 @@ def start_douyin_youtube_ffplay(
cleanup_proc_fn=cleanup_proc,
stderr_reader_fn=_stderr_reader_thread,
vocal_mix_ratio=demucs_vocal_mix,
mix_source=demucs_mix_source,
stats_interval_sec=stats_interval_sec,
)
except StreamEnded: