From 87bc1f8782d3c35228bd1ef5cfeefd1484ea1f41 Mon Sep 17 00:00:00 2001 From: zwa73 Date: Fri, 31 Mar 2023 01:29:02 +0800 Subject: [PATCH 1/4] =?UTF-8?q?=E6=A0=B9=E6=8D=AE@ChrisPreston=E7=9A=84?= =?UTF-8?q?=E4=BB=A3=E7=A0=81=E6=B7=BB=E5=8A=A0=E4=BA=86=E5=99=AA=E9=9F=B3?= =?UTF-8?q?=E8=BF=87=E6=BB=A4f0=5Ffilter=E5=8F=82=E6=95=B0=20=E5=9C=A8flas?= =?UTF-8?q?k=5Fapi=E4=B8=AD=E6=B7=BB=E5=8A=A0=E4=BA=86=E4=B8=80=E4=BA=9B?= =?UTF-8?q?=E5=8F=82=E6=95=B0=E4=BC=A0=E9=80=92?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- flask_api.py | 10 +++++++--- inference/infer_tool.py | 44 ++++++++++++++++++++++++++++++++++------- 2 files changed, 44 insertions(+), 10 deletions(-) diff --git a/flask_api.py b/flask_api.py index 8cc236a..5be4141 100644 --- a/flask_api.py +++ b/flask_api.py @@ -30,10 +30,13 @@ def voice_change_model(): # 模型推理 if raw_infer: - out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) + # out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) + out_audio, _ = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0, + auto_predict_f0=False, noice_scale=0.4, f0_filter=False) tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample) else: - out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path) + out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0, + auto_predict_f0=False, noice_scale=0.4, f0_filter=False) tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample) # 返回音频 out_wav_path = io.BytesIO() @@ -50,7 +53,8 @@ if __name__ == '__main__': # 每个模型和config是唯一对应的 model_name = "logs/32k/G_174000-Copy1.pth" config_name = "configs/config.json" - svc_model = Svc(model_name, config_name) + cluster_model_path = "logs/44k/kmeans_10000.pt" + svc_model = Svc(model_name, config_name, cluster_model_path=cluster_model_path) svc = RealTimeVC() # 此处与vst插件对应，不建议更改 app.run(port=6842, host="0.0.0.0", debug=False, threaded=False) diff --git a/inference/infer_tool.py b/inference/infer_tool.py index 3a2635b..b60c00f 100644 --- a/inference/infer_tool.py +++ b/inference/infer_tool.py @@ -104,6 +104,9 @@ def pad_array(arr, target_length): return padded_arr +class F0FilterException(Exception): + pass + class Svc(object): def __init__(self, net_g_path, config_path, device=None, @@ -138,11 +141,15 @@ class Svc(object): - def get_unit_f0(self, in_path, tran, cluster_infer_ratio, speaker): + def get_unit_f0(self, in_path, tran, cluster_infer_ratio, speaker, f0_filter): wav, sr = librosa.load(in_path, sr=self.target_sample) f0 = utils.compute_f0_parselmouth(wav, sampling_rate=self.target_sample, hop_length=self.hop_size) + + if f0_filter and sum(f0) == 0: + raise F0FilterException("未检测到人声") + f0, uv = utils.interpolate_f0(f0) f0 = torch.FloatTensor(f0) uv = torch.FloatTensor(uv) @@ -166,13 +173,14 @@ class Svc(object): def infer(self, speaker, tran, raw_path, cluster_infer_ratio=0, auto_predict_f0=False, - noice_scale=0.4): + noice_scale=0.4, + f0_filter=False): speaker_id = self.spk2id.__dict__.get(speaker) if not speaker_id and type(speaker) is int: if len(self.spk2id.__dict__) >= speaker: speaker_id = speaker sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0) - c, f0, uv = self.get_unit_f0(raw_path, tran, cluster_infer_ratio, speaker) + c, f0, uv = self.get_unit_f0(raw_path, tran, cluster_infer_ratio, speaker, f0_filter) if "half" in self.net_g_path and torch.cuda.is_available(): c = c.half() with torch.no_grad(): @@ -181,7 +189,7 @@ class Svc(object): use_time = time.time() - start print("vits use time:{}".format(use_time)) return audio, audio.shape[-1] - + def clear_empty(self): # 清理显存 torch.cuda.empty_cache() @@ -227,14 +235,30 @@ class RealTimeVC: """输入输出都是1维numpy 音频波形数组""" - def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path): + def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path, + cluster_infer_ratio=0, + auto_predict_f0=False, + noice_scale=0.4, + f0_filter=False): + + cluster_infer_ratio = cluster_infer_ratio + auto_predict_f0 = auto_predict_f0 + noice_scale = noice_scale + f0_filter = f0_filter + import maad audio, sr = torchaudio.load(input_wav_path) audio = audio.cpu().numpy()[0] temp_wav = io.BytesIO() if self.last_chunk is None: input_wav_path.seek(0) - audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) + + audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, + cluster_infer_ratio=cluster_infer_ratio, + auto_predict_f0=auto_predict_f0, + noice_scale=noice_scale, + f0_filter=f0_filter) + audio = audio.cpu().numpy() self.last_chunk = audio[-self.pre_len:] self.last_o = audio @@ -243,7 +267,13 @@ class RealTimeVC: audio = np.concatenate([self.last_chunk, audio]) soundfile.write(temp_wav, audio, sr, format="wav") temp_wav.seek(0) - audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav) + + audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav, + cluster_infer_ratio=cluster_infer_ratio, + auto_predict_f0=auto_predict_f0, + noice_scale=noice_scale, + f0_filter=f0_filter) + audio = audio.cpu().numpy() ret = maad.util.crossfade(self.last_o, audio, self.pre_len) self.last_chunk = audio[-self.pre_len:] From 061d3b46a5b16673613d54416924d3ac1ba1d81f Mon Sep 17 00:00:00 2001 From: zwa73 Date: Fri, 31 Mar 2023 01:34:10 +0800 Subject: [PATCH 2/4] fix --- inference/infer_tool.py | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/inference/infer_tool.py b/inference/infer_tool.py index 2da5c58..0ea8397 100644 --- a/inference/infer_tool.py +++ b/inference/infer_tool.py @@ -179,6 +179,7 @@ class Svc(object): auto_predict_f0=False, noice_scale=0.4, f0_filter=False): + speaker_id = self.spk2id.__dict__.get(speaker) if not speaker_id and type(speaker) is int: if len(self.spk2id.__dict__) >= speaker: @@ -266,11 +267,6 @@ class RealTimeVC: noice_scale=0.4, f0_filter=False): - cluster_infer_ratio = cluster_infer_ratio - auto_predict_f0 = auto_predict_f0 - noice_scale = noice_scale - f0_filter = f0_filter - import maad audio, sr = torchaudio.load(input_wav_path) audio = audio.cpu().numpy()[0] From 6a0000823299314204861aeb9da0a63fa4188791 Mon Sep 17 00:00:00 2001 From: zwa73 Date: Fri, 31 Mar 2023 01:35:50 +0800 Subject: [PATCH 3/4] fix --- flask_api.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/flask_api.py b/flask_api.py index 5be4141..5e5aebe 100644 --- a/flask_api.py +++ b/flask_api.py @@ -31,7 +31,7 @@ def voice_change_model(): # 模型推理 if raw_infer: # out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) - out_audio, _ = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0, + out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0, auto_predict_f0=False, noice_scale=0.4, f0_filter=False) tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample) else: From 41dd39790ff211b1bfbbf34ea168fb0f5301dea4 Mon Sep 17 00:00:00 2001 From: zwa73 Date: Fri, 31 Mar 2023 01:37:23 +0800 Subject: [PATCH 4/4] fix --- flask_api.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/flask_api.py b/flask_api.py index 5e5aebe..b3f1e06 100644 --- a/flask_api.py +++ b/flask_api.py @@ -32,7 +32,7 @@ def voice_change_model(): if raw_infer: # out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0, - auto_predict_f0=False, noice_scale=0.4, f0_filter=False) + auto_predict_f0=False, noice_scale=0.4, f0_filter=False) tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample) else: out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0,