Hass/xiaozhi-esp32-server
1
0
Fork 0
mirror of https://github.com/xinnan-tech/xiaozhi-esp32-server.git synced 2026-07-21 22:53:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #3285 from xinnan-tech/py-aec

Browse Source 
perf: 优化相关逻辑减少重复计算
This commit is contained in:
wengzh
2026-07-14 16:36:16 +08:00
committed by GitHub
parent ad3fb4d8c8 55c0495661
commit 1998807b69
1 changed files with 27 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files
main/xiaozhi-server/core/connection.py
+21 -46
View File
@@ -414,8 +414,6 @@ class ConnectionHandler:
def _apply_aec(self, timestamp: int, pcm_frame: bytes) -> bytes:
"""应用AEC处理 - 综合算法：互相关延迟估计 + Wiener滤波 + 频谱减法"""
try:
import numpy as np
if not pcm_frame or len(pcm_frame) == 0:
return pcm_frame
@@ -428,41 +426,27 @@ class ConnectionHandler:
if mic_rms < 100:
return pcm_frame
# 初始化
if not hasattr(self, "_aec_filter_state"):
self._aec_filter_state = np.zeros(512) # 滤波器状态
self._aec_delay_ms = 200
self._aec_last_delay = 200
self._aec_coherence_history = []
sorted_timestamps = sorted(self.aec_audio_cache.keys())
if len(sorted_timestamps) < 2:
return pcm_frame
# ========== 匹配参考帧（对数功率谱匹配） ==========
n = len(mic_audio)
sorted_timestamps = sorted(self.aec_audio_cache.keys())
# 找最接近的timestamp作为起点
closest_idx = min(range(len(sorted_timestamps)), key=lambda i: abs(sorted_timestamps[i] - timestamp))
# 对数功率谱相关性计算
def log_powerSpectrum_corr(audio1, audio2):
window = np.hanning(len(audio1))
fft1 = np.fft.rfft(audio1 * window)
fft2 = np.fft.rfft(audio2 * window)
psd1 = np.abs(fft1) ** 2
psd2 = np.abs(fft2) ** 2
log1 = 10 * np.log10(psd1 + 1e-8)
log2 = 10 * np.log10(psd2 + 1e-8)
P_xy = np.dot(log1, log2)
P_xx = np.dot(log1, log1)
P_yy = np.dot(log2, log2)
return abs(P_xy) / (np.sqrt(P_xx) * np.sqrt(P_yy) + 1e-8)
# 预计算 mic_audio 的对数功率谱（循环内共用，避免重复FFT）
mic_window = np.hanning(n)
mic_fft = np.fft.rfft(mic_audio * mic_window)
mic_psd = np.abs(mic_fft) ** 2
mic_log_psd = 10 * np.log10(mic_psd + 1e-8)
mic_P_xx = np.dot(mic_log_psd, mic_log_psd)
# 用对数功率谱匹配找最佳帧：前后各找2帧
best_corr = -1
best_ref_idx = closest_idx
best_ref_rms = 0.0
for offset in range(-2, 3): # T-2, T-1, T, T+1, T+2
test_idx = closest_idx + offset
@@ -475,16 +459,22 @@ class ConnectionHandler:
continue
# 对数功率谱相关性
corr = log_powerSpectrum_corr(mic_audio, test_ref)
test_window = np.hanning(len(test_ref))
test_fft = np.fft.rfft(test_ref * test_window)
test_psd = np.abs(test_fft) ** 2
test_log_psd = 10 * np.log10(test_psd + 1e-8)
P_xy = np.dot(mic_log_psd, test_log_psd)
P_yy = np.dot(test_log_psd, test_log_psd)
corr = abs(P_xy) / (np.sqrt(mic_P_xx) * np.sqrt(P_yy) + 1e-8)
if corr > best_corr:
best_corr = corr
best_ref_idx = test_idx
best_ref_rms = test_ref_rms
best_ts = sorted_timestamps[best_ref_idx]
best_ref = np.frombuffer(self.aec_audio_cache[best_ts], dtype=np.int16).astype(np.float32)
ref_rms = np.sqrt(np.mean(best_ref ** 2))
ref_rms = best_ref_rms
if ref_rms < 50:
return pcm_frame
@@ -499,44 +489,29 @@ class ConnectionHandler:
# 频域幅度谱不受相位影响，对数功率谱相关性稳定在0.97+
# 公式：result_mag = max(|mic_fft| - |ref_fft| * scale * coef, 0)
window = np.hanning(n)
mic_fft = np.fft.rfft(mic_audio * window)
ref_fft = np.fft.rfft(aligned_ref * window)
mic_mag = np.abs(mic_fft)
ref_mag = np.abs(ref_fft)
mic_phase = np.angle(mic_fft)
ref_fft = np.fft.rfft(aligned_ref * np.hanning(n))
ref_mag = np.abs(ref_fft)
# 频域计算回声比例 scale
scale = np.sum(mic_mag * ref_mag) / (np.dot(ref_mag, ref_mag) + 1e-8)
# 使用对数功率谱相关性作为coh
coh = best_corr
ref_power = ref_rms
# 自适应系数：根据scale和coh动态调整
# scale大（回声强）-> coef大；coh高（匹配准）-> coef大
raw_coef = 1.0 + scale * 3 + (coh - 0.97) * 30
raw_coef = 1.0 + scale * 3 + (best_corr - 0.97) * 30
coef = max(0.5, min(3.0, raw_coef))
# 谱减法（过减 + 半波整流）
if ref_power < 50:
output = mic_audio
coef = 0
gain = 0
else:
echo_mag = ref_mag * scale * coef
# 过减：使用较大系数抑制回声
over_subtract = 1.5
result_mag = np.maximum(mic_mag - echo_mag * over_subtract, mic_mag * 0.1)
result_mag = np.maximum(mic_mag - echo_mag * 1.5, mic_mag * 0.1)
# 保留相位重建信号
result_fft = result_mag * np.exp(1j * mic_phase)
output = np.fft.irfft(result_fft, n)
gain = scale * coef
# 高置信度是纯回声时，再压一下确保VAD检测不到
if coh >= 0.97 and ref_power > 500:
if best_corr >= 0.97 and ref_rms > 500:
output = output * 0.3
# 后处理：限幅