5.png

Llama-3B-Mono-Cooper

Llama-3B-Mono-Cooper 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 male voice named 'Cooper' 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-Cooper')
model = AutoModelForCausalLM.from_pretrained(
    'prithivMLmods/Llama-3B-Mono-Cooper', torch_dtype=torch.bfloat16
).cuda()

def generate_audio(text, temperature, top_p, max_new_tokens):
    """
    Given input text, generate speech audio.
    """
    speaker = "Cooper"
    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-Cooper - Single Speaker Audio Generation")
    gr.Markdown("Generate speech audio using the `prithivMLmods/Llama-3B-Mono-Cooper` 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-Cooper')
model = AutoModelForCausalLM.from_pretrained(
    'prithivMLmods/Llama-3B-Mono-Cooper', 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-Cooper - Single Speaker Audio Generation")
    gr.Markdown("Generate speech audio using the `prithivMLmods/Llama-3B-Mono-Cooper` 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.
Downloads last month
10
Safetensors
Model size
3.3B params
Tensor type
FP16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for prithivMLmods/Llama-3B-Mono-Cooper

Finetuned
(9)
this model
Quantizations
1 model

Collection including prithivMLmods/Llama-3B-Mono-Cooper