tikhonova

ai.train('result.txt',
         line_by_line=False,
         from_cache=False,
         num_steps=100000,
         generate_every=5000,
         save_every=5000,
         save_gdrive=True,
         learning_rate=1e-3,
         fp16=False,
         batch_size=1, 

tikhonova

''' ffmpeg not found when using pydub utils
___
If using Win, need to download from the official website and add to path, then reload git bash.
https://github.com/jiaaro/pydub/issues/348
'''

''' AssertionError: Distributed mode requires CUDA
___
a MUST-read to confirm that both GPU and drivers support the CUDA version you've installed (or about to install):
https://stackoverflow.com/questions/60987997/why-torch-cuda-is-available-returns-false-even-after-installing-pytorch-with

tikhonova

# Attention parameters
attention_rnn_dim=1024,  # sets the number of units in the RNN
attention_dim=128,  # sets the number of units in the attention mechanism
#  These two values are relatively large and may require a significant amount of GPU memory during training and inference.

# Location Layer parameters
attention_location_n_filters=32,  # sets the number of filters in the CNN
attention_location_kernel_size=31,  # sets the size of the filters
# This means that the CNN has 32 filters and each filter has a kernel size of 31.

tikhonova

''' Make metadata.csv and filelists via https://jaimeleal.github.io/how-to-speech-synthesis '''

import os
import pandas as pd
from sklearn.model_selection import train_test_split
import numpy as np

filepath = 'E:/AlanWatts/dataset/transcripts2/'
files = os.listdir(filepath)
rows = []

tikhonova

# snippet of instantiating a client
client = speech.SpeechClient()
config = speech.RecognitionConfig(
    encoding=speech.RecognitionConfig.AudioEncoding.LINEAR16,
    sample_rate_hertz=22050,
    audio_channel_count=1,
    model='phone_call',  # recognizes low quality audio better than default
    use_enhanced=1,  # if available
    language_code="en-US")
    

tikhonova

# using AudioSegment

    def get_duration(self):
        return self.audio.duration_seconds

    def single_split(self, from_min, to_min, split_filename):
        t1 = from_min * 7 * 1000  # convert to milliseconds
        t2 = to_min * 7 * 1000
        split_audio = self.audio[t1:t2]
        resampled = split_audio.set_frame_rate(22050)  # change the sampling rate to 22050 Hz

tikhonova

def remove_sil(file_path: str, file: str, dest_path: str, format="wav"):
    sound = AudioSegment.from_file(os.path.join(file_path, file), format=format)
    non_sil_times = detect_nonsilent(sound, min_silence_len=50, silence_thresh=sound.dBFS * 1.5)
    if len(non_sil_times) == 0:
        return None
    elif len(non_sil_times) > 0:
        non_sil_times_concat = [non_sil_times[0]]
        if len(non_sil_times) > 1:
            for t in non_sil_times[1:]:
                if t[0] - non_sil_times_concat[-1][-1] < 200:

tikhonova

def reduce_noise(file_path: str, file: str, dest_path: str) -> None:
    # load data
    rate, data = wavfile.read(os.path.join(file_path, file))

    reduced_noise = nr.reduce_noise(y=data, sr=rate)
    # perform noise reduction
    try:
        wavfile.write(os.path.join(dest_path, file), rate, reduced_noise)
    except Exception():
        pass

tikhonova

def convert_audio_to_wav(filename: str, filepath: str, dest_path: str) -> None:
    filepath = os.path.join(filepath, filename)
    dest_filepath = os.path.join(dest_path, f"{filename[:-4]}.wav")
    given_audio = AudioSegment.from_file(filepath, format="mp3")  # replace with mp4 or avi
    given_audio.export(dest_filepath, format="wav")

# create a list of input arguments for the function
inputs = [(filename, filepath, dest_path) for filename in os.listdir(filepath) if filename not in os.listdir(dest_path)]

# create a Process pool with 16 worker processes

tikhonova

# United States of America Python Dictionary to translate States,
# Districts & Territories to Two-Letter codes and vice versa.
#
# https://gist.github.com/rogerallen/1583593
#
# Dedicated to the public domain.  To the extent possible under law,
# Roger Allen has waived all copyright and related or neighboring
# rights to this code.

us_state_abbrev = {