janky way to generate more than 13 seconds

[privet2023]

Hello! I'm not a programmer, I made this using bing chat, but anyways!
This code splits given text into sequence of prompts, processes them one by one and stitches them back together
I guess it's possible to make it run faster and not (load model into vram??? or what the first progress bar does😨) every time
It also should fix CAPS and some punctuation of pasted text, I noticed it's messing up the result

from bark import SAMPLE_RATE, generate_audio, preload_models
from IPython.display import Audio
import numpy as np
import re

# download and load all models
preload_models()

# i forgot what it does i guess forces text to be in one single line bc otherwise it would error my jupyter thing

def format_text(text: str) -> str:
    lines = text.split('\n')
    formatted_lines = []
    for line in lines:
        if not line.endswith('.') and not line.endswith(','):
            line += '.'
        formatted_lines.append(line)
    formatted_prompt = ' '.join(formatted_lines)
    return formatted_prompt

# paste your text here

text = """your text here!!!"""

formatted_prompt = format_text(text)

# fix caps
def fix_caps(text):
    words = text.split()
    fixed_words = [word.capitalize() if word.isupper() else word for word in words]
    return ' '.join(fixed_words)

formatted_prompt = fix_caps(formatted_prompt)


# fix unnececary punctuation

# удаляем пробелы перед знаками препинания (кроме кавычек)
formatted_prompt = re.sub(r'\s+([^"\w\s])', r'\1', formatted_prompt)

# заменяем все знаки препинания, кроме первого, на пробел (кроме кавычек)
formatted_prompt = re.sub(r'([^\w\s"\'])([^\w\s"\']+)', r'\1 ', formatted_prompt)

# удаляем все лишние пробелы
formatted_prompt = re.sub(r'\s+', ' ', formatted_prompt).strip()

# generate audio from text
text_prompt = formatted_prompt
max_length = 140

def split_text(text, max_length):
    if len(text) <= max_length:
        return [text.lstrip()]
    else:
        split_index = max_length
        for delimiter in ['.', '!', '?', '♪', ')', ']', '}', '"', "'"]:
            index = text.rfind(delimiter, 0, max_length)
            if index != -1:
                split_index = min(split_index, index)
        if split_index == max_length:
            split_index = text.rfind(' ', 0, max_length)
        if split_index == -1:
            split_index = max_length

        # check if last character of the line is a quote and the next character is a period
        if text[split_index] == '"' or text[split_index] == "'":
            next_char_index = split_index + 1
            if next_char_index < len(text) and text[next_char_index] == '.':
                split_index += 1

        return [text[:split_index+1].lstrip()] + split_text(text[split_index+1:].lstrip(), max_length)

text_prompts = split_text(text_prompt, max_length)


# display separated prompts
for i, prompt in enumerate(text_prompts):
    print(f"Prompt {i+1}: {prompt}")
audio_arrays = [generate_audio(text, history_prompt="pl_speaker_4") for text in text_prompts]

# concatenate audio arrays into one
concatenated_audio = np.concatenate(audio_arrays)

# play concatenated audio in notebook
Audio(concatenated_audio, rate=SAMPLE_RATE)

[gkucsko]

there is also a PR where people are discussing different ways of doing this: #84