app.py

# !git clone https://github.com/Edresson/Coqui-TTS -b multilingual-torchaudio-SE TTS

from TTS.utils.manage import ModelManager
from TTS.utils.synthesizer import Synthesizer

manager = ModelManager()
model_path1, config_path1, model_item = manager.download_model("tts_models/zh-CN/baker/tacotron2-DDC-GST")
synthesizer = Synthesizer(
    model_path1, config_path1, None, None, None,
)

import os
import shutil
import gradio as gr

import sys

import string
import time
import argparse
import json

import numpy as np
# import IPython
# from IPython.display import Audio

import torch

from TTS.tts.utils.synthesis import synthesis
from TTS.tts.utils.text.symbols import make_symbols, phonemes, symbols
try:
  from TTS.utils.audio import AudioProcessor
except:
  from TTS.utils.audio import AudioProcessor


from TTS.tts.models import setup_model
from TTS.config import load_config
from TTS.tts.models.vits import *

from TTS.tts.utils.speakers import SpeakerManager
from pydub import AudioSegment

# from google.colab import files
import librosa

from scipy.io.wavfile import write, read

import subprocess

import openai

mes = [
    {"role": "system", "content": "You are my personal assistant. Try to be helpful. Respond to me only in Chinese."}
]


'''
from google.colab import drive
drive.mount('/content/drive')

src_path = os.path.join(os.path.join(os.path.join(os.path.join(os.getcwd(), 'drive'), 'MyDrive'), 'Colab Notebooks'), 'best_model_latest.pth.tar')
dst_path = os.path.join(os.getcwd(), 'best_model.pth.tar')

shutil.copy(src_path, dst_path)
'''

TTS_PATH = "TTS/"

# add libraries into environment
sys.path.append(TTS_PATH) # set this if TTS is not installed globally

# Paths definition

OUT_PATH = 'out/'

# create output path
os.makedirs(OUT_PATH, exist_ok=True)

# model vars 
MODEL_PATH = 'best_model.pth.tar'
CONFIG_PATH = 'config.json'
TTS_LANGUAGES = "language_ids.json"
TTS_SPEAKERS = "speakers.json"
USE_CUDA = torch.cuda.is_available()

# load the config
C = load_config(CONFIG_PATH)

# load the audio processor
ap = AudioProcessor(**C.audio)

speaker_embedding = None

C.model_args['d_vector_file'] = TTS_SPEAKERS
C.model_args['use_speaker_encoder_as_loss'] = False

model = setup_model(C)
model.language_manager.set_language_ids_from_file(TTS_LANGUAGES)
# print(model.language_manager.num_languages, model.embedded_language_dim)
# print(model.emb_l)
cp = torch.load(MODEL_PATH, map_location=torch.device('cpu'))
# remove speaker encoder
model_weights = cp['model'].copy()
for key in list(model_weights.keys()):
  if "speaker_encoder" in key:
    del model_weights[key]

model.load_state_dict(model_weights)

model.eval()

if USE_CUDA:
    model = model.cuda()

# synthesize voice
use_griffin_lim = False

# Paths definition

CONFIG_SE_PATH = "config_se.json"
CHECKPOINT_SE_PATH = "SE_checkpoint.pth.tar"

# Load the Speaker encoder

SE_speaker_manager = SpeakerManager(encoder_model_path=CHECKPOINT_SE_PATH, encoder_config_path=CONFIG_SE_PATH, use_cuda=USE_CUDA)

# Define helper function


def chatgpt(apikey, result):
    
    openai.api_key = apikey

    messages = mes

    # chatgpt
    content = result
    messages.append({"role": "user", "content": content})

    completion = openai.ChatCompletion.create(
      model = "gpt-3.5-turbo",
      messages = messages
    )

    chat_response = completion.choices[0].message.content

    messages.append({"role": "assistant", "content": chat_response}) 

    wavs = synthesizer.tts(chat_response + "。")
    
    synthesizer.save_wav(wavs, "output.wav")

    a1, b1 = read("output.wav")

    audio_out = "audio_out.wav"

    write(audio_out, a1, b1)

    return [chat_response, audio_out]

def compute_spec(ref_file):
  y, sr = librosa.load(ref_file, sr=ap.sample_rate)
  spec = ap.spectrogram(y)
  spec = torch.FloatTensor(spec).unsqueeze(0)
  return spec


def voice_conversion(ta, ra, da):

  target_audio = 'target.wav'
  reference_audio = 'reference.wav'
  driving_audio = 'driving.wav'

  write(target_audio, ta[0], ta[1])
  write(reference_audio, ra[0], ra[1])
  write(driving_audio, da[0], da[1])          

  # !ffmpeg-normalize $target_audio -nt rms -t=-27 -o $target_audio -ar 16000 -f
  # !ffmpeg-normalize $reference_audio -nt rms -t=-27 -o $reference_audio -ar 16000 -f
  # !ffmpeg-normalize $driving_audio -nt rms -t=-27 -o $driving_audio -ar 16000 -f

  files = [target_audio, reference_audio, driving_audio]

  for file in files:
      subprocess.run(["ffmpeg-normalize", file, "-nt", "rms", "-t=-27", "-o", file, "-ar", "16000", "-f"])

  # ta_ = read(target_audio)

  target_emb = SE_speaker_manager.compute_d_vector_from_clip([target_audio])
  target_emb = torch.FloatTensor(target_emb).unsqueeze(0)

  driving_emb = SE_speaker_manager.compute_d_vector_from_clip([reference_audio])
  driving_emb = torch.FloatTensor(driving_emb).unsqueeze(0)

  # Convert the voice

  driving_spec = compute_spec(driving_audio)
  y_lengths = torch.tensor([driving_spec.size(-1)])
  if USE_CUDA:
      ref_wav_voc, _, _ = model.voice_conversion(driving_spec.cuda(), y_lengths.cuda(), driving_emb.cuda(), target_emb.cuda())
      ref_wav_voc = ref_wav_voc.squeeze().cpu().detach().numpy()
  else:
      ref_wav_voc, _, _ = model.voice_conversion(driving_spec, y_lengths, driving_emb, target_emb)
      ref_wav_voc = ref_wav_voc.squeeze().detach().numpy()

  # print("Reference Audio after decoder:")
  # IPython.display.display(Audio(ref_wav_voc, rate=ap.sample_rate))

  return (ap.sample_rate, ref_wav_voc)

    
block = gr.Blocks()

with block:
    with gr.Group():
        gr.Markdown(
            """ # <center>🥳💬💕 - TalktoAI，随时随地，谈天说地！</center>
            
            ## <center>🤖 - 让有人文关怀的AI造福每一个人！AI向善，文明璀璨！TalktoAI - Enable the future！</center>
            
      """
        )
        
        with gr.Box():
            with gr.Row().style(mobile_collapse=False, equal_height=True):

                inp1 = gr.components.Textbox(lines=2, label="请填写您的OpenAI-API-key")
                inp2 = gr.components.Textbox(lines=2, label="说些什么吧")

                btn = gr.Button("开始对话吧")
 
        texts = gr.Textbox(lines=2, label="ChatGPT的回答")
        audio_tts = gr.Audio(label="自动合成的声音")
              
        btn.click(chatgpt, [inp1, inp2], [texts, audio_tts])

        with gr.Box():
            with gr.Row().style(mobile_collapse=False, equal_height=True):
                inp3 = gr.Audio(label = "请上传您喜欢的声音(wav/mp3文件, max. 30mb)")
                inp4 = gr.Audio(label = "被转换的语音")
                inp5 = gr.Audio(label = "被转换的语音")

                btn1 = gr.Button("用喜欢的声音听一听吧")

        out1 = gr.Audio(label="声音拟合的专属声音")

        btn1.click(voice_conversion, [inp3, inp4, inp5], [out1])

        gr.Markdown(
            """ 
            
            ### <center>注意❗：请不要输入或生成会对个人以及组织造成侵害的内容，此程序仅供科研、学习及娱乐使用。用户输入或生成的内容与程序开发者无关，请自觉合法合规使用，违反者一切后果自负。</center>
            
            ### <center>Model by [Raven](https://huggingface.co/spaces/BlinkDL/Raven-RWKV-7B). Thanks to [PENG Bo](https://github.com/BlinkDL). Please follow me on [Bilibili](https://space.bilibili.com/501495851?spm_id_from=333.1007.0.0).</center>
            
      """
        )
        
        gr.HTML('''
        <div class="footer">
                    <p>🎶🖼️🎡 - It’s the intersection of technology and liberal arts that makes our hearts sing. - Steve Jobs
                    </p>
        </div>
        ''')


block.launch(show_error=True)