Llama-3B-Mono-Ceylia

Llama-3B-Mono-Ceylia is a Llama-based Speech-LLM designed for high-quality, empathetic text-to-speech generation. This model has been fine-tuned to deliver human-like speech synthesis, achieving exceptional clarity, expressiveness, and real-time streaming performance. The model has been fine-tuned from mono audio of a female voice named 'Ceylia' using the base model canopylabs/orpheus-3b-0.1-ft.

In some cases, the results may be inconsistent, particularly when handling complex speech transformations.

[ paralinguistic emotions soft]

Model Details

• Base Model: canopylabs/orpheus-3b-0.1-ft
• Languages Supported: English
• License: Llama 3.2
• Model Version: N/A

---

Paralinguistic Elements

The model can generate speech with the following emotions:

Elements	Elements	Elements
laugh	chuckle	sigh
sniffle	groan	yawn
gasp	uhm	giggles & more

---

Run with Transformers 🤗

from huggingface_hub import notebook_login, HfApi
notebook_login()

Install Dependencies

%%capture
!pip install snac accelerate
!pip install transformers
!pip install gradio

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
from snac import SNAC

def redistribute_codes(row):
    """
    Convert a sequence of token codes into an audio waveform using SNAC.
    The code assumes each 7 tokens represent one group of instructions.
    """
    row_length = row.size(0)
    new_length = (row_length // 7) * 7
    trimmed_row = row[:new_length]
    code_list = [t - 128266 for t in trimmed_row]
    
    layer_1, layer_2, layer_3 = [], [], []
    
    for i in range((len(code_list) + 1) // 7):
        layer_1.append(code_list[7 * i][None])
        layer_2.append(code_list[7 * i + 1][None] - 4096)
        layer_3.append(code_list[7 * i + 2][None] - (2 * 4096))
        layer_3.append(code_list[7 * i + 3][None] - (3 * 4096))
        layer_2.append(code_list[7 * i + 4][None] - (4 * 4096))
        layer_3.append(code_list[7 * i + 5][None] - (5 * 4096))
        layer_3.append(code_list[7 * i + 6][None] - (6 * 4096))
    
    with torch.no_grad():
        codes = [
            torch.concat(layer_1),
            torch.concat(layer_2),
            torch.concat(layer_3)
        ]
        for i in range(len(codes)):
            codes[i][codes[i] < 0] = 0
            codes[i] = codes[i][None]
        
        audio_hat = snac_model.decode(codes)
        return audio_hat.cpu()[0, 0]

# Load the SNAC model for audio decoding
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to("cuda")

# Load the single-speaker language model
tokenizer = AutoTokenizer.from_pretrained('prithivMLmods/Llama-3B-Mono-Ceylia')
model = AutoModelForCausalLM.from_pretrained(
    'prithivMLmods/Llama-3B-Mono-Ceylia', torch_dtype=torch.bfloat16
).cuda()

def generate_audio(text, temperature, top_p, max_new_tokens):
    """
    Given input text, generate speech audio.
    """
    speaker = "Ceylia"
    prompt = f'<custom_token_3><|begin_of_text|>{speaker}: {text}<|eot_id|><custom_token_4><custom_token_5><custom_token_1>'
    input_ids = tokenizer(prompt, add_special_tokens=False, return_tensors='pt').to('cuda')
    
    with torch.no_grad():
        generated_ids = model.generate(
            **input_ids,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=temperature,
            top_p=top_p,
            repetition_penalty=1.1,
            num_return_sequences=1,
            eos_token_id=128258,
        )
    
    row = generated_ids[0, input_ids['input_ids'].shape[1]:]
    y_tensor = redistribute_codes(row)
    y_np = y_tensor.detach().cpu().numpy()
    return (24000, y_np)

# Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("# Llama-3B-Mono-Ceylia - Single Speaker Audio Generation")
    gr.Markdown("Generate speech audio using the `prithivMLmods/Llama-3B-Mono-Ceylia` model.")
    
    with gr.Row():
        text_input = gr.Textbox(lines=4, label="Input Text")
    
    with gr.Row():
        temp_slider = gr.Slider(minimum=0.1, maximum=2.0, step=0.1, value=0.9, label="Temperature")
        top_p_slider = gr.Slider(minimum=0.1, maximum=1.0, step=0.05, value=0.8, label="Top-p")
        tokens_slider = gr.Slider(minimum=100, maximum=2000, step=50, value=1200, label="Max New Tokens")
    
    output_audio = gr.Audio(type="numpy", label="Generated Audio")
    generate_button = gr.Button("Generate Audio")
    
    generate_button.click(
        fn=generate_audio,
        inputs=[text_input, temp_slider, top_p_slider, tokens_slider],
        outputs=output_audio
    )

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

[ or ]

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
from snac import SNAC

def redistribute_codes(row):
    """
    Convert a sequence of token codes into an audio waveform using SNAC.
    The code assumes each 7 tokens represent one group of instructions.
    """
    row_length = row.size(0)
    new_length = (row_length // 7) * 7
    trimmed_row = row[:new_length]
    code_list = [t - 128266 for t in trimmed_row]
    
    layer_1, layer_2, layer_3 = [], [], []
    
    for i in range((len(code_list) + 1) // 7):
        layer_1.append(code_list[7 * i][None])
        layer_2.append(code_list[7 * i + 1][None] - 4096)
        layer_3.append(code_list[7 * i + 2][None] - (2 * 4096))
        layer_3.append(code_list[7 * i + 3][None] - (3 * 4096))
        layer_2.append(code_list[7 * i + 4][None] - (4 * 4096))
        layer_3.append(code_list[7 * i + 5][None] - (5 * 4096))
        layer_3.append(code_list[7 * i + 6][None] - (6 * 4096))
    
    with torch.no_grad():
        codes = [
            torch.concat(layer_1),
            torch.concat(layer_2),
            torch.concat(layer_3)
        ]
        for i in range(len(codes)):
            codes[i][codes[i] < 0] = 0
            codes[i] = codes[i][None]
        
        audio_hat = snac_model.decode(codes)
        return audio_hat.cpu()[0, 0]

# Load the SNAC model for audio decoding
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to("cuda")

# Load the single-speaker language model
tokenizer = AutoTokenizer.from_pretrained('prithivMLmods/Llama-3B-Mono-Ceylia')
model = AutoModelForCausalLM.from_pretrained(
    'prithivMLmods/Llama-3B-Mono-Ceylia', torch_dtype=torch.bfloat16
).cuda()

def generate_audio(text, temperature, top_p, max_new_tokens):
    """
    Given input text, generate speech audio.
    """
    prompt = f'<custom_token_3><|begin_of_text|>{text}<|eot_id|><custom_token_4><custom_token_5><custom_token_1>'
    input_ids = tokenizer(prompt, add_special_tokens=False, return_tensors='pt').to('cuda')
    
    with torch.no_grad():
        generated_ids = model.generate(
            **input_ids,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=temperature,
            top_p=top_p,
            repetition_penalty=1.1,
            num_return_sequences=1,
            eos_token_id=128258,
        )
    
    row = generated_ids[0, input_ids['input_ids'].shape[1]:]
    y_tensor = redistribute_codes(row)
    y_np = y_tensor.detach().cpu().numpy()
    return (24000, y_np)

# Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("# Llama-3B-Mono-Ceylia - Single Speaker Audio Generation")
    gr.Markdown("Generate speech audio using the `prithivMLmods/Llama-3B-Mono-Ceylia` model.")
    
    with gr.Row():
        text_input = gr.Textbox(lines=4, label="Input Text")
    
    with gr.Row():
        temp_slider = gr.Slider(minimum=0.1, maximum=2.0, step=0.1, value=0.9, label="Temperature")
        top_p_slider = gr.Slider(minimum=0.1, maximum=1.0, step=0.05, value=0.8, label="Top-p")
        tokens_slider = gr.Slider(minimum=100, maximum=2000, step=50, value=1200, label="Max New Tokens")
    
    output_audio = gr.Audio(type="numpy", label="Generated Audio")
    generate_button = gr.Button("Generate Audio")
    
    generate_button.click(
        fn=generate_audio,
        inputs=[text_input, temp_slider, top_p_slider, tokens_slider],
        outputs=output_audio
    )

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

---

Intended Use

• Designed for high-quality, single-speaker text-to-speech generation.
• Ideal for applications requiring human-like speech synthesis.
• Supports a range of emotions for expressive speech output.
• Suitable for AI voice assistants, storytelling, and accessibility applications.