somrajmondal/PrettyVoice_model

PrettyVoice – Expressive Text-to-Speech Model

Model Overview

PrettyVoice is an expressive Text-to-Speech (TTS) and Chat-to-Speech model fine-tuned to generate emotional, conversational, and natural human-like speech.

This model is optimized for acted dialogue, emotional expression, and conversational delivery, not flat or robotic narration.

⚠️ This is an audio-generation model, not a text-only language model.

---

Model Details

Field	Value
Model name	somrajmondal/PrettyVoice_model
Author	Somraj Mondal
Architecture	CsmForConditionalGeneration
Language	English
Output	Speech audio (WAV, 24 kHz)
Speaker support	Single speaker (speaker_id = 0)
License	Apache 2.0

---

Intended Use

✅ Designed for

• Text-to-Speech (TTS)
• Chat-style speech generation
• Emotional and expressive dialogue
• Storytelling and voice acting
• Romantic and conversational AI voices
• Multimodal chat → audio pipelines

❌ Not designed for

• Text-only generation
• Code generation
• Speech-to-text (ASR)
• Multilingual speech
• Robotic or factual narration

---

Installation

pip install unsloth transformers soundfile torch accelerate safetensors

Inference Code

import torch
import soundfile as sf
from IPython.display import Audio, display
from unsloth import FastModel
from transformers import CsmForConditionalGeneration

# Load model and processor
model, processor = FastModel.from_pretrained(
    model_name="somrajmondal/PrettyVoice_model",
    auto_model=CsmForConditionalGeneration,
    max_seq_length=2048,
)

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



from IPython.display import Audio, display
import soundfile as sf
import torch

text = '''<smiles> Hey guys! How are you today? How sweet does my voice sound?'''
speaker_id = 0
conversation = [
    {"role": str(speaker_id), "content": [{"type": "text", "text": text}]},
]

# -------------------------
# Prepare inputs and compute token length
# -------------------------
inputs_for_length = processor.apply_chat_template(
    conversation,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
)

# Number of input tokens
input_length = inputs_for_length["input_ids"].shape[1]

# Estimate max_new_tokens for audio generation
max_new_tokens = int(input_length * 2.5)
# Clamp min and max
max_new_tokens = max(150, min(max_new_tokens, 3600))

print(f"Input tokens: {input_length}, max_new_tokens set to: {max_new_tokens}")

# -------------------------
# Generate audio
# -------------------------
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

audio_values = model.generate(
    **inputs_for_length.to(device),
    max_new_tokens=max_new_tokens,
    output_audio=True,
    # do_sample=True,       # makes voice expressive
    # temperature=0.9,
    # top_p=0.9,
)

# -------------------------
# Convert to numpy, normalize, save and play
# -------------------------
audio = audio_values[0].float().cpu().numpy()
audio = audio / (abs(audio).max() + 1e-6)  # normalize

sf.write("example_voice_context.wav", audio, 24000)
display(Audio(audio, rate=24000))

print("Saved example_voice_context.wav successfully!")