import gradio as gr from transformers import AutoProcessor, BarkModel import torch import scipy # Limit CPU usage torch.set_num_threads(1) # Load the Bark model and processor processor = AutoProcessor.from_pretrained("suno/bark-small") model = BarkModel.from_pretrained("suno/bark-small") # Function to generate speech def generate_speech(text, voice_preset): # Process the input text with the selected voice preset inputs = processor(text, voice_preset=voice_preset) # Generate audio and convert to float32 early to optimize memory usage with torch.no_grad(): # Disable gradient calculations for faster inference audio_array = model.generate(**inputs) audio_array = audio_array.cpu().numpy().astype('float32').squeeze() # Converting early # Return the audio with sample rate for Gradio's audio component return (model.generation_config.sample_rate, audio_array) # Gradio app setup with gr.Blocks() as app: gr.Markdown("# Turkish Text-to-Speech with Bark") gr.Markdown("Enter text, select a Turkish voice preset, and click 'Generate Voice' to play the generated audio.") # Input text box for user to type text text_input = gr.Textbox(label="Enter Text in Turkish", placeholder="Merhaba, bugün bir yerlere gidelim mi?") # Dropdown for selecting voice preset voice_preset_input = gr.Dropdown( ["v2/tr_speaker_0", "v2/tr_speaker_1", "v2/tr_speaker_2", "v2/tr_speaker_3", "v2/tr_speaker_4", "v2/tr_speaker_5", "v2/tr_speaker_6", "v2/tr_speaker_7", "v2/tr_speaker_8", "v2/tr_speaker_9"], label="Select Turkish Voice Preset" ) # Audio output component for playing generated audio audio_output = gr.Audio(label="Generated Voice", type="numpy") # Button to trigger the generation generate_button = gr.Button("Generate Voice") # When the button is clicked, call the generate_speech function generate_button.click(generate_speech, inputs=[text_input, voice_preset_input], outputs=audio_output) # Launch the Gradio app app.launch()