别盲目喂音频做音色克隆：千条人声素材标准化预处理流水线实战

pip install librosa soundfile pydub webrtcvad tqdm numpy

import os
import librosa
import numpy as np
import soundfile as sf
from pydub import AudioSegment
import webrtcvad
from tqdm import tqdm
# =====================配置参数，根据你的模型修改=====================
SAMPLE_RATE = 16000
MIN_TARGET_SEC = 10
MAX_TARGET_SEC = 30
SNR_THRESHOLD = 12  #信噪比低于该值直接丢弃
RAW_AUDIO_DIR = "./1000voice_raw"    #网盘解压原始素材文件夹
CLEAN_OUTPUT_DIR = "./voice_clean"   #处理完成合格音频输出
# =================================================================
os.makedirs(CLEAN_OUTPUT_DIR, exist_ok=True)
def calculate_snr(signal, noise_floor_db=-40):
    """简易信噪比计算"""
    power = np.sum(signal ** 2) / len(signal)
    noise_power = 10 ** (noise_floor_db / 10)
    if power <= noise_power:
        return 0
    snr = 10 * np.log10(power / noise_power)
    return snr
def convert_audio(file_path):
    """统一转为16k单声道numpy数组"""
    y, sr = librosa.load(file_path, sr=SAMPLE_RATE, mono=True)
    return y
def slice_by_vad(wave_data):
    """webrtcvad做语音活动检测，切分有效人声片段"""
    vad = webrtcvad.Vad(2)
    frame_ms = 30
    samples_per_frame = int(SAMPLE_RATE * frame_ms / 1000)
    segments = []
    temp_segment = []
    for i in range(0, len(wave_data), samples_per_frame):
        frame = wave_data[i:i+samples_per_frame]
        if len(frame) < samples_per_frame:
            break
        frame_pcm = (frame * 32767).astype(np.int16).tobytes()
        is_speech = vad.is_speech(frame_pcm, SAMPLE_RATE)
        if is_speech:
            temp_segment.extend(frame)
        else:
            if len(temp_segment) > 0:
                segments.append(np.array(temp_segment))
                temp_segment = []
    if len(temp_segment) > 0:
        segments.append(np.array(temp_segment))
    return segments
def norm_loudness(wave_data, target_db=-18):
    """响度归一化"""
    audio_seg = AudioSegment(
        data=(wave_data*32767).astype(np.int16).tobytes(),
        sample_width=2, frame_rate=SAMPLE_RATE, channels=1
    )
    change_db = target_db - audio_seg.dBFS
    normalized = audio_seg.apply_gain(change_db)
    arr = np.array(normalized.get_array_of_samples()) / 32767.0
    return arr.astype(np.float32)
def process_single_audio(file_name):
    full_path = os.path.join(RAW_AUDIO_DIR, file_name)
    try:
        wave = convert_audio(full_path)
        seg_list = slice_by_vad(wave)
        idx = 0
        for seg in seg_list:
            duration = len(seg)/SAMPLE_RATE
            if not (MIN_TARGET_SEC <= duration <= MAX_TARGET_SEC):
                continue
            snr = calculate_snr(seg)
            if snr < SNR_THRESHOLD:
                continue
            seg_norm = norm_loudness(seg)
            out_name = f"{os.path.splitext(file_name)[0]}_{idx}.wav"
            out_path = os.path.join(CLEAN_OUTPUT_DIR, out_name)
            sf.write(out_path, seg_norm, SAMPLE_RATE)
            idx += 1
    except Exception as e:
        print(f"跳过异常文件 {file_name} : {str(e)}")
if __name__ == "__main__":
    file_list = [f for f in os.listdir(RAW_AUDIO_DIR) if f.lower().endswith((".wav",".mp3",".m4a"))]
    for fname in tqdm(file_list, desc="素材预处理"):
        process_single_audio(fname)
    print(f"\n✅预处理完成，合格音频输出目录：{CLEAN_OUTPUT_DIR}")

from TTS.api import TTS
USE_GPU = True
#选取预处理完成的一条音色参考音频
REF_AUDIO = "./voice_clean/sample_0.wav"
OUT_WAV = "./drama_output/role_speech.wav"
text_content = "夜色慢慢笼罩城市，那些埋藏心底的秘密，终于要浮出水面。"
if __name__ == "__main__":
    tts = TTS(model_name="tts_models/multilingual/multi-dataset/xtts_v2", gpu=USE_GPU)
    tts.tts_to_file(
        text=text_content,
        speaker_wav=REF_AUDIO,
        language="zh-cn",
        file_path=OUT_WAV
    )
    print(f"✅克隆配音输出至 {OUT_WAV}")