app.py

import gradio as gr
import torch
from transformers import BarkModel
from optimum.bettertransformer import BetterTransformer

model = BarkModel.from_pretrained("suno/bark", torch_dtype=torch.float16)
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = model.to(device)

from transformers import AutoProcessor
processor = AutoProcessor.from_pretrained("suno/bark")

# Use bettertransform for flash attention
model = BetterTransformer.transform(model, keep_original_model=False)

# Enable CPU offload
model.enable_cpu_offload()

import numpy as np
from scipy.io.wavfile import write as write_wav
import wave

def split_text_into_sentences(text):
    sentences = []
    current_sentence = ''
    words = text.split()

    for word in words:
        current_sentence += ' ' + word
        if word.endswith('.'):
            sentences.append(current_sentence.strip())
            current_sentence = ''

    if current_sentence:
        sentences.append(current_sentence.strip())

    return sentences

def join_wav_files(input_files, output_file):
    # Open the first input file to get its parameters
    with wave.open(input_files[0], 'rb') as first_file:
        # Get the audio parameters from the first file
        params = first_file.getparams()

        # Create a new wave file for writing the joined audio
        with wave.open(output_file, 'wb') as output:
            output.setparams(params)

            # Iterate over the input files and write their audio data to the output file
            for input_file in input_files:
                with wave.open(input_file, 'rb') as input:
                    output.writeframes(input.readframes(input.getnframes()))


def infer(text_prompt):
    print("""
    —
    Cutting text in chunks
    —
    """)
    
    
    text_chunks = split_text_into_sentences(text_prompt)
   
    result = generate(text_chunks, "wav")
    print(result)
    
    
    output_wav = 'full_story.wav'

    join_wav_files(result, output_wav)

    return 'full_story.wav'


def generate(text_prompt, out_type):
    text_prompt = text_prompt

    inputs = processor(text_prompt, voice_preset="v2/en_speaker_6").to(device)
    
    with torch.inference_mode():
        speech_output = model.generate(**inputs)

    input_waves = []
    
    for i, speech_out in enumerate(speech_output):
        
        audio_array = speech_out.cpu().numpy().squeeze()
        print(f'AUDIO_ARRAY: {audio_array}')
    
        # Assuming audio_array contains audio data and the sampling rate
        sampling_rate = model.generation_config.sample_rate
        print(f'sampling_rate: {sampling_rate}')

        if out_type == "numpy":
            input_waves.append(sampling_rate, audio_array)
        elif out_type == "wav":
            #If you want to return a WAV file :
            # Ensure the audio data is properly scaled (between -1 and 1 for 16-bit audio)
            
            audio_data = np.int16(audio_array * 32767)  # Scale for 16-bit signed integer
            write_wav(f"output_{i}.wav", sampling_rate, audio_data)
            input_waves.append(f"output_{i}.wav")
    return input_waves


with gr.Blocks() as demo:
    with gr.Column():
        prompt = gr.Textbox(label="prompt")
        submit_btn = gr.Button("Submit")
        audio_out = gr.Audio()
    submit_btn.click(fn=infer, inputs=[prompt], outputs=[audio_out])

demo.launch()