Skip to content

Commit 9588a3c

Browse files
RVC-BossRVC-Boss
authored
support speed adjust支持合成语速调整
support speed adjust支持合成语速调整
1 parent 5fbe7f0 commit 9588a3c

1 file changed

Lines changed: 45 additions & 42 deletions

File tree

‎GPT_SoVITS/inference_webui.py‎

Lines changed: 45 additions & 42 deletions
Original file line numberDiff line numberDiff line change
@@ -312,7 +312,11 @@ def merge_short_text_in_array(texts, threshold):
312312
result[len(result) - 1] += text
313313
return result
314314

315-
def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False,speed=1):
315+
##ref_wav_path+prompt_text+prompt_language+text(单个)+text_language+top_k+top_p+temperature
316+
# cache_tokens={}#暂未实现清理机制
317+
cache=None
318+
def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False,speed=1,if_freeze=False):
319+
global cache
316320
if prompt_text is None or len(prompt_text) == 0:
317321
ref_free = True
318322
t0 = ttime()
@@ -395,38 +399,30 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
395399
all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
396400

397401
t2 = ttime()
398-
with torch.no_grad():
399-
# pred_semantic = t2s_model.model.infer(
400-
pred_semantic, idx = t2s_model.model.infer_panel(
401-
all_phoneme_ids,
402-
all_phoneme_len,
403-
None if ref_free else prompt,
404-
bert,
405-
# prompt_phone_len=ph_offset,
406-
top_k=top_k,
407-
top_p=top_p,
408-
temperature=temperature,
409-
early_stop_num=hz * max_sec,
410-
)
402+
# cache_key="%s-%s-%s-%s-%s-%s-%s-%s"%(ref_wav_path,prompt_text,prompt_language,text,text_language,top_k,top_p,temperature)
403+
if(type(cache)!=type(None)and if_freeze==True):pred_semantic=cache
404+
else:
405+
with torch.no_grad():
406+
pred_semantic, idx = t2s_model.model.infer_panel(
407+
all_phoneme_ids,
408+
all_phoneme_len,
409+
None if ref_free else prompt,
410+
bert,
411+
# prompt_phone_len=ph_offset,
412+
top_k=top_k,
413+
top_p=top_p,
414+
temperature=temperature,
415+
early_stop_num=hz * max_sec,
416+
)
417+
pred_semantic = pred_semantic[:, -idx:].unsqueeze(0)
418+
cache=pred_semantic
411419
t3 = ttime()
412-
# print(pred_semantic.shape,idx)
413-
pred_semantic = pred_semantic[:, -idx:].unsqueeze(
414-
0
415-
) # .unsqueeze(0)#mq要多unsqueeze一次
416420
refer = get_spepc(hps, ref_wav_path) # .to(device)
417421
if is_half == True:
418422
refer = refer.half().to(device)
419423
else:
420424
refer = refer.to(device)
421-
# audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
422-
audio = (
423-
vq_model.decode(
424-
pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer,speed=speed
425-
)
426-
.detach()
427-
.cpu()
428-
.numpy()[0, 0]
429-
) ###试试重建不带上prompt部分
425+
audio = (vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer,speed=speed).detach().cpu().numpy()[0, 0])
430426
max_audio=np.abs(audio).max()#简单防止16bit爆音
431427
if max_audio>1:audio/=max_audio
432428
audio_opt.append(audio)
@@ -611,29 +607,36 @@ def get_weights_names():
611607
)
612608
gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式"))
613609
with gr.Row():
614-
text = gr.Textbox(label=i18n("需要合成的文本"), value="")
615-
text_language = gr.Dropdown(
616-
label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文")
617-
)
618-
how_to_cut = gr.Radio(
619-
label=i18n("怎么切"),
620-
choices=[i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ],
621-
value=i18n("凑四句一切"),
622-
interactive=True,
623-
)
624-
with gr.Row():
625-
gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低)："))
610+
with gr.Column():
611+
text = gr.Textbox(label=i18n("需要合成的文本"), value="")
612+
text_language = gr.Dropdown(
613+
label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文")
614+
)
615+
how_to_cut = gr.Radio(
616+
label=i18n("怎么切"),
617+
choices=[i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ],
618+
value=i18n("凑四句一切"),
619+
interactive=True,
620+
)
621+
with gr.Column():
622+
gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低。不懂就用默认)："))
626623
top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=10,interactive=True)
627624
top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True)
628625
temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True)
629-
with gr.Row():
630-
speed = gr.Slider(minimum=0.5,maximum=2,step=0.05,label=i18n("speed"),value=1,interactive=True)
626+
with gr.Column():
627+
gr.Markdown(value=i18n("语速调整，高为更快"))
628+
if_freeze=gr.Checkbox(label=i18n("是否直接对上次合成结果调整语速。防止随机性。"), value=False, interactive=True, show_label=True)
629+
speed = gr.Slider(minimum=0.6,maximum=1.65,step=0.05,label=i18n("语速"),value=1,interactive=True)
630+
# with gr.Column():
631+
# gr.Markdown(value=i18n("手工调整音素。当音素框不为空时使用手工音素输入推理，无视目标文本框。"))
632+
# phoneme=gr.Textbox(label=i18n("音素框"), value="")
633+
# get_phoneme_button = gr.Button(i18n("目标文本转音素"), variant="primary")
631634
inference_button = gr.Button(i18n("合成语音"), variant="primary")
632635
output = gr.Audio(label=i18n("输出的语音"))
633636

634637
inference_button.click(
635638
get_tts_wav,
636-
[inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free,speed],
639+
[inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free,speed,if_freeze],
637640
[output],
638641
)
639642

0 commit comments

Comments
 (0)