语音分离模型：mossfomer2
- 计算短时能量
def compute_short_time_energy(audio: np.ndarray, frame_size: int, hop_size: int) -> np.ndarray:"""计算音频信号的短时能量+ 将音频分为若干帧，每一帧长度为 frame_size, 帧与帧之间以 hop_size 的步长滑动。+ 每一帧的能量定义为该帧所有采样点的平方和。Args:audio (array): 1D音频序列frame_size (int): 帧长: 多少秒内的采样点数hop_size (int): 帧移: 多少秒内的采样点数Returns:array: 能量数组，每个元素对应一帧的能量"""num_frames = int(np.ceil((len(audio) - frame_size) / hop_size)) + 1energy = np.zeros(num_frames)for i in range(num_frames):start = i * hop_sizeend = min(start + frame_size, len(audio))frame = audio[start:end]energy[i] = np.sum(frame ** 2)return energy
- 主声源提取
def segment_audio(audio, sr, energy, threshold):"""利用短时能量定位音频中的主声源区域Args:audio (array): 1D音频序列sr (int): 采样率energy (array): 每帧的短时能量数组threshold (float): 能量阈值，低于该值认为振幅较低Returns:list: 列表，每个元组包含 (开始时间, 结束时间, 主声源片段, 短时能量值)"""segments = []acc_energy = []in_segment = Falsestart_frame = Nonelast_active_frame = None# 遍历每一帧能量，检测连续区域for i, e in enumerate(energy):if e >= threshold:# 当前帧能量达到阈值，若尚未进入段则记录起始帧if not in_segment:in_segment = Truestart_frame = ilast_active_frame = i # 更新最后一次达到阈值的帧acc_energy.append(e)else:# 当前帧能量低于阈值且之前处于主声源区域，完成一次分段if in_segment:start_sample = start_frame * hop_size# 以最后一次能量达到阈值的帧末尾作为结束位置end_sample = min(last_active_frame * hop_size + frame_size, len(audio))segments.append((start_sample / sr, end_sample / sr, audio[start_sample:end_sample], np.mean(acc_energy)))in_segment = Falsestart_frame = Nonelast_active_frame = None# 若音频结束时仍在主声源区域，则补全该段if in_segment:start_sample = start_frame * hop_sizeend_sample = len(audio)segments.append((start_sample / sr, end_sample / sr, audio[start_sample:end_sample]))return segmentsdef save_segments(segments, sr, output_dir="segments"):"""保存检测到的主声源片段为独立的音频文件，并合并保存所有片段Args:segments (list): 分段结果列表，每个元组为 (开始时间, 结束时间, 主声源片段)sr (int): 采样率output_dir (str, optional): 保存的目录（默认保存到 "segments" 文件夹）"""if not os.path.exists(output_dir):os.makedirs(output_dir)merged_audio_list = [seg[-2] for seg in segments]# 合并所有片段并保存为一个文件if merged_audio_list:merged_audio = np.concatenate(merged_audio_list)merged_file = os.path.join(output_dir, "merged_segments.wav")sf.write(merged_file, merged_audio, sr)print(f"已保存合并文件：{merged_file}")
# audio_file = "/home/wangguisen/projects/voice_separation/output/单轨音频/普通单轨音频_5_spk0.wav"audio_file = "/home/wangguisen/projects/voice_separation/2-.wav"audio, sr = sf.read(audio_file)# 参数设置：例如帧长20ms，帧移10msframe_size = int(0.02 * sr) # 20ms 对应的采样点数hop_size = int(0.01 * sr) # 10ms 对应的采样点数print("[DEBUG] frame_size: ", frame_size)print("[DEBUG] hop_size: ", hop_size)# 设置能量阈值（根据实际情况调整），低于此值认为振幅较低threshold = 0.001# 计算短时能量energy = compute_short_time_energy(audio, frame_size, hop_size)# 利用短时能量分段定位主声源segments = segment_audio(audio, sr, energy, threshold)# 输出检测到的主声源区间print("检测到的主声源区间：")for seg in segments:start, end, _, e = segprint(f"开始时间: {start:.3f} s, 结束时间: {end:.3f} s, 短时能量值: {e: .3f}")# 保存每段主声源到独立文件，并合并保存所有片段save_segments(segments, sr, output_dir="segments")
