Dash Text to Speech

speech_synthesis_app.py

# -*- coding: utf-8 -*-
"""
Speech Synthesis with Dash. This will only work with a Nvidia GPU.

Requirements (put them in requirements.txt):

apex
dash
numpy
scipy
torch
"""
import base64
import time
import io

from apex import amp
import dash
import dash_core_components as dcc
import dash_html_components as html
from dash.dependencies import Input, Output, State
import numpy as np
from scipy.io.wavfile import write
import torch


# Load model
tacotron2 = torch.hub.load('nvidia/DeepLearningExamples:torchhub', 'nvidia_tacotron2')
tacotron2 = tacotron2.to('cuda')
tacotron2 = amp.initialize(tacotron2, opt_level="O1")
tacotron2.eval()

waveglow = torch.hub.load('nvidia/DeepLearningExamples:torchhub', 'nvidia_waveglow')
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow = waveglow.to('cuda')
waveglow = amp.initialize(waveglow, opt_level="O1")
waveglow.eval()


# Dash app starts here
external_stylesheets = ['https://codepen.io/chriddyp/pen/bWLwgP.css']

app = dash.Dash(__name__, external_stylesheets=external_stylesheets)

app.layout = html.Div(children=[
    html.H1(children='Dash Text-to-Speech'),

    dcc.Textarea(
        id='textarea-input',
        value=text,
        style={'width': '100%', 'height': '45vh'}
    ),
    
    dcc.Loading([
        html.Button("Generate", id='button'),
        html.Audio(id='audio-out', controls=True)
    ])
])


@app.callback(Output("audio-out", "src"),
              [Input("button", "n_clicks")],
              [State("textarea-input", "value")])
def generate_audio(n_clicks, text):
    if text == "":
        text = "Sorry, there's nothing in the text input. Please write something."
        
    t0 = time.time()
    # preprocessing
    sequence = np.array(tacotron2.text_to_sequence(text, ['english_cleaners']))[None, :]
    sequence = torch.from_numpy(sequence).to(device='cuda', dtype=torch.int64)

    # run the models
    with torch.no_grad():
        _, mel, _, _ = tacotron2.infer(sequence)
        audio = waveglow.infer(mel)

    audio_numpy = audio[0].data.cpu().numpy()
    rate = 22050

    t1 = time.time()

    buffer = io.BytesIO()
    write(buffer, rate, audio_numpy)
    b64 = base64.b64encode(buffer.getvalue())
    sound = "data:audio/x-wav;base64," + b64.decode("ascii")

    t2 = time.time()

    print(f"Completed in {t2-t0:.3f}s. Generation took {t1-t0:.3f}s, file creation took {t2-t1:.3f}s")

    return sound

if __name__ == "__main__":
    app.run_server(debug=False)