chatTTS系列之-webui优化

背景

ChatTTS官方默认webui可供选择功能太少,我们今天来改造一下

在此基础上在做一些改造诸如音色保存等功能,并对功能实现做一下记录。有错误还请评论区勘正。

增加控制参数

增加speed、oral、laugh、break等参数

# 添加oral_slider  laugh_slider  break_slider 三个滑块
with gr.Row():
      speed_slider = gr.Slider(minimum=0, maximum=9, step=1, value=2, label="speed")
      oral_slider = gr.Slider(minimum=0, maximum=9, step=1, value=3, label="oral")
      laugh_slider = gr.Slider(minimum=0, maximum=9, step=1, value=3, label="laugh")
      break_slider = gr.Slider(minimum=0, maximum=9, step=1, value=3, label="break")

这样在启用优化文本时就会添加诸如笑声、停顿等

params_refine_text = {'prompt': f'[oral_{refine_oral}][laugh_{refine_laugh}][break_{refine_break}]'}
    
if refine_text_flag:
        text = chat.infer(text, 
                          ...
                          params_refine_text=params_refine_text,
                          )

保存音色

绘制界面

with gr.Row():
    text_input = gr.Textbox(label="音色名称", lines=4, placeholder="请输入音色名称", value="")
    save_btn = gr.Button("保存音色")

将当前声音tensor参数存储下来

info = {}
info['spk_emb'] = ''
spk_embs_dir = Path("spk_embs")

if not spk_embs_dir.exists():
    spk_embs_dir.mkdir()

...

rand_spk = chat.sample_random_speaker()
info['spk_emb'] = rand_spk


...
def save_voice(voice_name):
    try:
        spk_emb = info['spk_emb']
        file_path = spk_embs_dir.joinpath(f'{voice_name}.json')
        if file_path.exists():
            return (f'{voice_name}.json already exists!')
        data_list = spk_emb.tolist()  
        with open(file_path, 'w') as f:
            json.dump(data_list, f)
        return f'{voice_name}.json saved!'
    except Exception as e:
        return f'{voice_name}.json save failed! {e}'

读取音色

绘制界面

with gr.Row():
    customer_speaker = gr.Checkbox(label="是否启用自定义音色", value=False)
    # 音色下拉框
    speaker = gr.Dropdown(choices=voice_infos.keys(), label="Speaker")

程序启动时加载音色

def read_voice_list():
    files = spk_embs_dir.glob("*.json")
    for file in files:
        try:
            with open(file, 'r', encoding='utf-8') as json_file:
                tensor = json.load(json_file)
            voice_infos[file.stem] = tensor
        except:
            print(f"Failed to read {file}")

read_voice_list()

生成音频时加载tensor

if customer_speaker_flag and speaker:
    if speaker not in voice_infos:
        return [(0, None), ""]
    speak_tensor = torch.tensor(voice_infos[speaker])
    params_infer_code = {"spk_emb": speak_tensor, "temperature": 0.0001}

...
wav = chat.infer(...
                         params_infer_code=params_infer_code
                         )

长文本生成

将长文本切分成段落,依次生成

paragraphs = re.split(r"\n+", text)

for paragraph in paragraphs:
    if refine_text_flag:
        paragraph = chat.infer(paragraph,
                               skip_refine_text=False,
                               refine_text_only=True,
                               params_refine_text=params_refine_text,
                               params_infer_code=params_infer_code
                               )

    text_data = paragraph[0] if isinstance(paragraph, list) else paragraph
    text_data_all.append(text_data)

    # 生成音频文件
    wav = chat.infer(paragraph,
                     skip_refine_text=True,
                     params_refine_text=params_refine_text,
                     params_infer_code=params_infer_code
                     )
    audio_data = np.array(wav[0]).flatten()
    audio_data_all.append(audio_data)

# 将所有的语音片段合并成一个完整的音频
audio_data_all = np.concatenate(audio_data_all)

自动保存当前音频

定义一个key保存当前音频存储路径

info = {}
info["current_voice"] = None

音频生成后自定保存到output文件夹

 # 检查通道数
 if audio_data_all.ndim == 1:  # 单声道
     num_channels = 1
 elif audio_data_all.ndim == 2:  # 立体声
     num_channels = audio_data_all.shape[1]

 audio_data_int16 = (audio_data_all * 32767).astype(np.int16)

 # 写入WAV文件
 # 将内存文件保存到磁盘上
 timestamp = datetime.now().strftime('%Y-%m-%d_%H_%M_%S')
 file_path = str(output.joinpath(f'{timestamp}.wav').resolve())
 info["current_voice"] = file_path
 wavfile.write(file_path, sample_rate, audio_data_all)

总结

做了一番改造后，功能丰富了不少，日常会易用很多。但仍不完美，等待后续做进一步改造。