AI 音色克隆数据集工程化方案：素材清洗、切片与质量筛选全链路

pip install librosa pydub webrtcvad numpy soundfile tqdm

import os
import librosa
import numpy as np
import soundfile as sf
from pydub import AudioSegment
import webrtcvad
from tqdm import tqdm
# ===================== 配置项（根据你的克隆模型调整） =====================
SAMPLE_RATE = 16000          # VAD标准采样率，大部分克隆模型兼容16k
TARGET_SEC_MIN = 12          # 输出片段最小时长
TARGET_SEC_MAX = 30          # 输出片段最大时长（克隆最优区间）
SNR_THRESHOLD = 12           # 信噪比阈值，低于该值直接丢弃
OUTPUT_DIR = "./clean_voice_ref"
RAW_DIR = "./raw_voice_data"
# =======================================================================
os.makedirs(OUTPUT_DIR, exist_ok=True)
def load_audio(file_path: str, target_sr: int = 16000):
    """加载音频，统一转单声道、目标采样率"""
    y, sr = librosa.load(file_path, sr=target_sr, mono=True)
    return y, target_sr
def calculate_snr(y: np.ndarray, sr: int):
    """简易信噪比估算，区分静音段与语音段"""
    energy = librosa.feature.rms(y=y)[0]
    noise_floor = np.percentile(energy, 10)
    speech_energy = np.percentile(energy, 70)
    if noise_floor <= 0:
        return 0.0
    snr = 10 * np.log10(speech_energy / noise_floor)
    return round(snr, 2)
def vad_split(y: np.ndarray, sr: int, frame_duration_ms=30, padding_ms=300):
    """WebRTC VAD 人声检测，返回有效语音时间戳列表"""
    vad = webrtcvad.Vad(2)
    bytes_per_frame = int(sr * frame_duration_ms / 1000 * 2)
    frames = []
    for i in range(0, len(y), int(sr * frame_duration_ms / 1000)):
        frame = y[i:i + int(sr * frame_duration_ms / 1000)]
        if len(frame) < int(sr * frame_duration_ms / 1000):
            break
        frame_int16 = (frame * 32767).astype(np.int16)
        is_speech = vad.is_speech(frame_int16.tobytes(), sr)
        frames.append(is_speech)
    # 合并连续语音块
    segments = []
    in_speech = False
    start = 0
    for idx, flag in enumerate(frames):
        if flag and not in_speech:
            start = idx * frame_duration_ms
            in_speech = True
        elif not flag and in_speech:
            end = idx * frame_duration_ms
            segments.append((start / 1000, end / 1000))
            in_speech = False
    return segments
def normalize_loudness(y: np.ndarray, target_db=-16):
    """峰值响度归一化，避免爆音、音量差异过大"""
    peak = np.max(np.abs(y))
    if peak < 1e-6:
        return y
    gain = 10 ** (target_db / 20) / peak
    return y * gain
def process_single_audio(file_path, out_dir):
    try:
        y, sr = load_audio(file_path, SAMPLE_RATE)
        snr = calculate_snr(y, sr)
        if snr < SNR_THRESHOLD:
            return False, f"信噪比不足 {snr}dB"
        speech_segs = vad_split(y, sr)
        seg_idx = 0
        for (t_start, t_end) in speech_segs:
            dur = t_end - t_start
            if dur < TARGET_SEC_MIN:
                continue
            # 长片段滑动切分
            for s in np.arange(t_start, t_end - TARGET_SEC_MIN, TARGET_SEC_MAX):
                e = min(s + TARGET_SEC_MAX, t_end)
                sub_y = y[int(s*sr): int(e*sr)]
                sub_y = normalize_loudness(sub_y)
                base_name = os.path.splitext(os.path.basename(file_path))[0]
                out_name = f"{base_name}_seg{seg_idx}.wav"
                sf.write(os.path.join(out_dir, out_name), sub_y, sr)
                seg_idx += 1
        return True, f"成功输出 {seg_idx} 个片段"
    except Exception as err:
        return False, str(err)
if __name__ == "__main__":
    audio_files = [f for f in os.listdir(RAW_DIR) if f.lower().endswith((".wav", ".mp3"))]
    for fname in tqdm(audio_files):
        full_path = os.path.join(RAW_DIR, fname)
        ok, msg = process_single_audio(full_path, OUTPUT_DIR)
        # print(f"[{fname}] {msg}")

from TTS.api import TTS
# 加载本地部署的TTS克隆模型
tts_model = TTS(model_name="本地语音复刻模型", progress_bar=False, gpu=True)
# 选择清洗后的干净参考片段
ref_audio = "./clean_voice_ref/demo_seg0.wav"
output_wav = "./gen_output.wav"
# 生成目标文本语音
tts_model.tts_to_file(
    text="今天我们来聊一下AI短剧配音的音色调优技巧。",
    file_path=output_wav,
    speaker_wav=ref_audio,
    language="zh"
)