app.py

import gradio as gr
import librosa
import numpy as np
import torch
import os
import torch
from speechbrain.pretrained import EncoderClassifier
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
from scipy.io import wavfile
import scipy.signal as sps
import openai as ai
import gc
from examples import *

checkpoint = "microsoft/speecht5_tts"
processor = SpeechT5Processor.from_pretrained(checkpoint)
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
ai.api_key = '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'


spk_model_name = "speechbrain/spkrec-xvect-voxceleb"

device = "cuda" if torch.cuda.is_available() else "cpu"
speaker_model = EncoderClassifier.from_hparams(
    source=spk_model_name,
    run_opts={"device": device},
    savedir=os.path.join("/tmp", spk_model_name))

def create_speaker_embedding(waveform):
    with torch.no_grad():
        speaker_embeddings = speaker_model.encode_batch(torch.tensor(waveform))
        speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
        speaker_embeddings = speaker_embeddings.squeeze().cpu().numpy()
    return speaker_embeddings

def prepare_data(temp_text, audio_prompt):
    rate, audio_data = audio_prompt
    # new_rate = 16000
    # number_of_samples = round(len(audio_data) * float(new_rate) / rate)
    # audio_data = sps.resample(audio_data, number_of_samples)
    example = processor(
        text=temp_text,
        audio_target=audio_data,
        sampling_rate=16000,
        return_attention_mask=False,)
    example["speaker_embeddings"] = create_speaker_embedding(audio_data)
    example_embeddings = torch.tensor(example["speaker_embeddings"]).unsqueeze(0)
    return example_embeddings

# def generate_gpt4_response(user_text, print_output=False):
#     """
#     Query OpenAI GPT-4 for the specific key and get back a response
#     :type user_text: str the user's text to query for
#     :type print_output: boolean whether or not to print the raw output JSON
#     """
#     message=[{"role": "user", "content": user_text+'in just 2 very small sentences'}]
#     completions = ai.ChatCompletion.create(
#          model="gpt-4",
#          messages=message,
#          max_tokens=250
#      )

#     # Return the first choice's text
#     return completions['choices'][0]['message']['content']


def predict(temp_text, temp_audio, record_audio_prompt, prompt_text):
    if temp_audio is not None :
      audio_prompt = temp_audio 
    else:
      audio_prompt = record_audio_prompt
    
    # text = generate_gpt4_response(prompt_text)
    text=prompt_text
    embeddings=prepare_data(temp_text, audio_prompt)
    inputs = processor(text=text, return_tensors="pt")
    spectrogram = model.generate_speech(inputs["input_ids"], embeddings)

    with torch.no_grad():
        speech = vocoder(spectrogram)

    speech = (speech.numpy() * 32767).astype(np.int16)
    speech=(16000, speech)
    del temp_text, temp_audio, record_audio_prompt, prompt_text, audio_prompt,embeddings,inputs,spectrogram
    gc.collect()
    return text, speech

app = gr.Blocks()
with app:
  with gr.Row():
    with gr.Column():

      temp_text=gr.Text(label="Template Text")
      temp_audio=gr.Audio(label="Template Speech", type="numpy")
      prompt_text=gr.Text(label="Input Text")
      record_audio_prompt = gr.Audio(label='recorded audio prompt', source='microphone', type="numpy")
    with gr.Column():
      text = gr.Textbox(label="Message")
      speech=gr.Audio(label="Generated Speech", type="numpy")
      btn = gr.Button("Generate!")
      btn.click(predict,
                inputs=[temp_text, temp_audio, record_audio_prompt, prompt_text],
                outputs=[text, speech])
  gr.Examples(examples=infer_from_audio_examples,
                inputs=[temp_text, temp_audio, record_audio_prompt, prompt_text],
                outputs=[text, speech],
                fn=predict,
                cache_examples=False,)

app.launch()