Quality and speed concerns/test for Turkish language

[emircanerkul]

Hi, I'm testing this model with Turkish i used one script showed in the docs, modified a bit to use gpu, but this wasn't good and each run it become worse.

from transformers import AutoProcessor, BarkModel
import gradio as gr
import torch
import pandas as pd

processor = AutoProcessor.from_pretrained("suno/bark")
model = BarkModel.from_pretrained("suno/bark")

# Move model to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

# Read voice presets from CSV
df = pd.read_csv('data.csv', header=None)
voice_presets = {f"{row[0]} ({row[2]})": row[1] for _, row in df.iterrows()}

def bark(text, voice_preset):
    # Process text with padding and attention mask
    inputs = processor(
        text, voice_preset
    )

    # Move inputs to GPU first
    inputs = {
        k: v.to(device) if hasattr(v, 'to') else v
        for k, v in inputs.items()
    }

    # Get attention mask from the moved inputs
    # attention_mask = inputs["attention_mask"]

    audio_array = model.generate(
        input_ids=inputs["input_ids"],
        pad_token_id=processor.tokenizer.pad_token_id
    )
    audio_array = audio_array.cpu().numpy().squeeze()
    sample_rate = model.generation_config.sample_rate
    return (sample_rate, audio_array)

interface = gr.Interface(
    fn=bark,
    inputs=[
        gr.Textbox(label="Text to speak", placeholder="Enter text here..."),
        gr.Dropdown(
            choices=list(voice_presets.items()),
            value=list(voice_presets.items())[0][1],  # Set first value as default
            label="Voice"
        )
    ],
    outputs=gr.Audio(label="Generated Speech"),
    title="Bark Text-to-Speech",
    description="Generate speech from text using different voices",
)

if __name__ == "__main__":
    interface.launch()

Then i cloned huggingface space and it was better but still far from good for turkish language also generation speed become much more slower like ~1m

What do you suggest me?

[emircanerkul]

Also, time to time,

• it changes the speaker
• put some different words end of the voice
• put a lot of silences (can be 10s silence)

With this version Turkish is not usable