长音频批量合成

.gitignore

@@ -13,5 +13,4 @@ processcmd.py
 .vscode
 WPy64-38100
 Winpython64-3.8.10.0dot.exe
-*.wav
 *.pkf

infer.py

@@ -1,4 +1,5 @@
 import logging
+from pathlib import Path
 
 import soundfile
 
@@ -13,8 +14,8 @@
 model_path = f'./checkpoints/{project_name}/model_ckpt_steps_44000.ckpt'
 
 # 支持多个wav文件，放在raw文件夹下
-clean_names = ["十年"]
-trans = [-6]  # 音高调整，支持正负（半音）
+file_names = ["mxj_61674.ogg"]
+trans = [-9]  # 音高调整，支持正负（半音）
 # 加速倍数
 accelerate = 50
 
@@ -26,16 +27,18 @@
 cut_time = 30
 
 svc_model = Svc(project_name, model_path)
-infer_tool.fill_a_to_b(trans, clean_names)
+infer_tool.fill_a_to_b(trans, file_names)
 infer_tool.mkdir(["./infer/wav_temp", input_wav_path, out_wav_path])
 
 # 清除缓存文件
 infer_tool.del_temp_wav(input_wav_path)
-for clean_name, tran in zip(clean_names, trans):
-    raw_audio_path = f"./raw/{clean_name}.wav"
+for f_name, tran in zip(file_names, trans):
+    raw_audio_path = f"./raw/{f_name}"
+    infer_tool.format_wav(raw_audio_path)
+    clean_name = f_name[:-4]
     infer_tool.del_temp_wav("./infer/wav_temp")
     out_audio_name = clean_name
-    infer_tool.cut_wav(raw_audio_path, out_audio_name, input_wav_path, cut_time)
+    infer_tool.cut_wav(Path(raw_audio_path).with_suffix('.wav'), out_audio_name, input_wav_path, cut_time)
 
     count = 0
     file_list = infer_tool.get_end_file(input_wav_path, "wav")
@@ -44,7 +47,8 @@
         raw_path = f"{input_wav_path}/{file_name}"
         out_path = f"{out_wav_path}/{file_name}"
 
-        audio = svc_model.infer(raw_path, key=tran, acc=accelerate, use_pe=True, use_gt_mel=False, add_noise_step=500)
+        f0_gt, f0_pred, audio = svc_model.infer(raw_path, key=tran, acc=accelerate, use_pe=True, use_gt_mel=False,
+                                                add_noise_step=500)
         soundfile.write(out_path, audio, 24000, 'PCM_16')
 
         count += 1

infer/infer_tool.py

@@ -4,7 +4,9 @@
 import subprocess
 import time
 
+import librosa
 import numpy as np
+import soundfile
 import torch
 import torchaudio
 
@@ -35,6 +37,11 @@ def run(*args, **kwargs):
     return run
 
 
+def format_wav(audio_path):
+    raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True)
+    soundfile.write(audio_path[:-4] + ".wav", raw_audio, raw_sample_rate)
+
+
 def cut_wav(raw_audio_path, out_audio_name, input_wav_path, cut_time):
     raw_audio, raw_sr = torchaudio.load(raw_audio_path)
     if raw_audio.shape[-1] / raw_sr > cut_time:
@@ -140,18 +147,33 @@ def after_infer(self, prediction):
                 prediction[k] = v.cpu().numpy()
 
         # remove paddings
+        mel_gt = prediction["mels"]
+        mel_gt_mask = np.abs(mel_gt).sum(-1) > 0
+        mel_gt = mel_gt[mel_gt_mask]
+        mel2ph_gt = prediction.get("mel2ph")
+        mel2ph_gt = mel2ph_gt[mel_gt_mask] if mel2ph_gt is not None else None
         mel_pred = prediction["outputs"]
         mel_pred_mask = np.abs(mel_pred).sum(-1) > 0
         mel_pred = mel_pred[mel_pred_mask]
+        mel_gt = np.clip(mel_gt, hparams['mel_vmin'], hparams['mel_vmax'])
         mel_pred = np.clip(mel_pred, hparams['mel_vmin'], hparams['mel_vmax'])
 
+        mel2ph_pred = prediction.get("mel2ph_pred")
+        if mel2ph_pred is not None:
+            if len(mel2ph_pred) > len(mel_pred_mask):
+                mel2ph_pred = mel2ph_pred[:len(mel_pred_mask)]
+            mel2ph_pred = mel2ph_pred[mel_pred_mask]
+
+        f0_gt = prediction.get("f0_gt")
         f0_pred = prediction.get("f0_pred")
+        if f0_pred is not None:
+            f0_gt = f0_gt[mel_gt_mask]
         if len(f0_pred) > len(mel_pred_mask):
             f0_pred = f0_pred[:len(mel_pred_mask)]
         f0_pred = f0_pred[mel_pred_mask]
         torch.cuda.is_available() and torch.cuda.empty_cache()
         wav_pred = self.vocoder.spec2wav(mel_pred, f0=f0_pred)
-        return wav_pred
+        return f0_gt, f0_pred, wav_pred
 
     def temporary_dict2processed_input(self, item_name, temp_dict, use_crepe=True, thre=0.05):
         '''