fix

.gitignore

@@ -0,0 +1 @@
+.idea/

app.py

@@ -1,20 +1,24 @@
+import os
+
+os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'
+
 from music.search import get_random_spit, get_albums
 from vits.models import SynthesizerInfer
+import whisper.inference
 from omegaconf import OmegaConf
 import torchcrepe
 import torch
 import io
-import os
 import gradio as gr
 import librosa
 import numpy as np
 import soundfile
 import random
-from audio2numpy import open_audio
 from spleeter.separator import Separator
 from spleeter.audio.adapter import AudioAdapter
 from pydub import AudioSegment
 import scipy.io.wavfile
+import uuid
 
 import logging
 
@@ -84,11 +88,13 @@ model.eval()
 model.to(device)
 separator = Separator('spleeter:2stems')
 audio_loader = AudioAdapter.default()
+whisper_model = whisper.inference.load_model(os.path.join("whisper_pretrain", "medium.pt"))
 
 
 def svc_change(argswave, argsspk):
     argsppg = "svc_tmp.ppg.npy"
-    os.system(f"python whisper/inference.py -w {argswave} -p {argsppg}")
+    whisper.inference.pred_ppg(whisper_model, argswave, argsppg)
+    # os.system(f"python whisper/inference.py -w {argswave} -p {argsppg}")
 
     spk = np.load(argsspk)
     spk = torch.FloatTensor(spk)
@@ -120,16 +126,16 @@ def svc_change(argswave, argsspk):
         out_audio = []
         has_audio = False
 
-        while (out_index + out_chunk < all_frame):
+        while out_index + out_chunk < all_frame:
             has_audio = True
-            if (out_index == 0):  # start frame
+            if out_index == 0:  # start frame
                 cut_s = out_index
                 cut_s_48k = 0
             else:
                 cut_s = out_index - hop_frame
                 cut_s_48k = hop_frame * hop_size
 
-            if (out_index + out_chunk + hop_frame > all_frame):  # end frame
+            if out_index + out_chunk + hop_frame > all_frame:  # end frame
                 cut_e = out_index + out_chunk
                 cut_e_48k = 0
             else:
@@ -148,8 +154,8 @@ def svc_change(argswave, argsspk):
             out_audio.extend(sub_out)
             out_index = out_index + out_chunk
 
-        if (out_index < all_frame):
-            if (has_audio):
+        if out_index < all_frame:
+            if has_audio:
                 cut_s = out_index - hop_frame
                 cut_s_48k = hop_frame * hop_size
             else:
@@ -177,23 +183,40 @@ def np_to_audio_segment(fp_arr):
     return sound
 
 
+def get_dtype_max_value(dtype):
+    if np.issubdtype(dtype, np.integer):
+        info = np.iinfo(dtype)
+        return info.max
+    elif np.issubdtype(dtype, np.floating):
+        info = np.finfo(dtype)
+        return info.max
+    else:
+        raise ValueError("不支持的 dtype 类型")
+
+
 def svc_main(sid, input_audio):
     if input_audio is None:
         return "You need to upload an audio", None
     sampling_rate, audio = input_audio
-    input_audio_tmp_file = 'origin.wav'
+    uuid_value = uuid.uuid4()
+    uuid_string = str(uuid_value)
+    input_audio_tmp_file = f'{uuid_string}.wav'
+    tmpfile_path = '/tmp'
+
     #
     # prediction = separator.separate(audio)
     # vocals, accompaniment = prediction["vocals"], prediction["accompaniment"]
     soundfile.write(input_audio_tmp_file, audio, sampling_rate, format="wav")
-    separator.separate_to_file(input_audio_tmp_file, '')
+    if not os.path.exists(tmpfile_path):
+        os.makedirs(tmpfile_path)
+    separator.separate_to_file(input_audio_tmp_file, tmpfile_path)
 
-    vocals_filepath = os.path.join(os.path.splitext(input_audio_tmp_file)[0], 'vocals.wav')
-    accompaniment_filepath = os.path.join(os.path.splitext(input_audio_tmp_file)[0], 'accompaniment.wav')
+    curr_tmp_path = os.path.join(tmpfile_path, os.path.splitext(input_audio_tmp_file)[0])
+    vocals_filepath = os.path.join(curr_tmp_path, 'vocals.wav')
+    accompaniment_filepath = os.path.join(curr_tmp_path, 'accompaniment.wav')
 
     vocals, sampling_rate = soundfile.read(vocals_filepath)
 
-    vocals = (vocals / np.iinfo(vocals.dtype).max).astype(np.float32)
     if len(vocals.shape) > 1:
         vocals = librosa.to_mono(vocals.transpose(1, 0))
     if sampling_rate != 16000:
@@ -204,7 +227,7 @@ def svc_main(sid, input_audio):
     soundfile.write(wav_path, vocals, 16000, format="wav")
 
     out_vocals = svc_change(wav_path, f"configs/singers/singer00{sid}.npy")
-    out_vocals_filepath = os.path.join(os.path.splitext(input_audio_tmp_file)[0], 'out_vocals.wav')
+    out_vocals_filepath = os.path.join(curr_tmp_path, 'out_vocals.wav')
     soundfile.write(out_vocals_filepath, out_vocals, 48000, format="wav")
 
     sound1 = AudioSegment.from_file(out_vocals_filepath)
@@ -212,7 +235,11 @@ def svc_main(sid, input_audio):
 
     played_togther = sound1.overlay(sound2)
 
-    return "Success", (48000, played_togther)
+    result_path = os.path.join(curr_tmp_path, 'out_song.wav')
+    played_togther.export(result_path, format="wav")
+    result, sampling_rate = soundfile.read(result_path)
+
+    return "Success", (sampling_rate, result)
 
 
 def auto_search(name):
@@ -221,7 +248,7 @@ def auto_search(name):
     album = random.choice(albums)
     save_path = get_random_spit(album)
     fp = save_path
-    signal, sampling_rate = open_audio(fp)
+    signal, sampling_rate = soundfile.read(fp)
     return sampling_rate, signal
 
 

requirements.txt

@@ -14,5 +14,4 @@ tqdm
 librosa
 pydub
 musicdl
-audio2numpy
 spleeter

whisper/inference.py

@@ -21,7 +21,7 @@ def pred_ppg(whisper: Whisper, wavPath, ppgPath):
     audln = audio.shape[0]
     ppg_a = []
     idx_s = 0
-    while (idx_s + 25 * 16000 < audln):
+    while idx_s + 25 * 16000 < audln:
         short = audio[idx_s:idx_s + 25 * 16000]
         idx_s = idx_s + 25 * 16000
         ppgln = 25 * 16000 // 320
@@ -31,7 +31,7 @@ def pred_ppg(whisper: Whisper, wavPath, ppgPath):
             ppg = whisper.encoder(mel.unsqueeze(0)).squeeze().data.cpu().float().numpy()
             ppg = ppg[:ppgln,]  # [length, dim=1024]
             ppg_a.extend(ppg)
-    if (idx_s < audln):
+    if idx_s < audln:
         short = audio[idx_s:audln]
         ppgln = (audln - idx_s) // 320
         # short = pad_or_trim(short)
@@ -48,6 +48,7 @@ if __name__ == "__main__":
     parser.description = 'please enter embed parameter ...'
     parser.add_argument("-w", "--wav", help="wav", dest="wav")
     parser.add_argument("-p", "--ppg", help="ppg", dest="ppg")
+    
     args = parser.parse_args()
     print(args.wav)
     print(args.ppg)