Orpheus Bangla Emotional TTS Model

This is a Text-to-Speech model for Bangla (Bengali) language based on the Orpheus architecture. The model can generate natural-sounding Bangla speech from text input with various emotional styles.

🎮 Demo

Try out the model directly using our Hugging Face Space:

Model Details

Model Type: Text-to-Speech
Language: Bangla (Bengali)
Architecture: Orpheus TTS + SNAC decoder
Sample Rate: 24kHz
Special Features: Emotional speech synthesis
Emotions: Various emotional tones including happy, sad, angry, disgusted, frustrated, excited, curious, surprised, etc.

Supported Emotions

This model supports the following emotional styles:

happy
normal
disgust
sad
frustrated
slow
excited
whisper
panicky
curious
surprise
fast
crying
deep
sleepy
angry
high
shout

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from snac import SNAC
import torch
import soundfile as sf

# Load the models
tokenizer = AutoTokenizer.from_pretrained("ehzawad/orpheus-bangla-emotional-tts")
model = AutoModelForCausalLM.from_pretrained(
    "ehzawad/orpheus-bangla-emotional-tts", 
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to("cuda")

# Sample prompt in Bangla
prompt = "আপনি কেমন আছেন?"  # "How are you?" in Bangla

# For emotional speech, add the emotion tag (e.g., for happy)
emotional_prompt = "<happy>আপনি কেমন আছেন?</happy>"

# Add your inference code here
# (Follow the example_usage.py code for complete inference)

Citation

If you use this model, please cite:

@misc{ehzawad2025orpheusbangla,
  author = {Ehzawad},
  title = {Orpheus Bangla Emotional TTS},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/ehzawad/orpheus-bangla-emotional-tts}}
}

Acknowledgements

This model is based on the Orpheus TTS architecture developed by Canopy Labs. We extend our gratitude to the original authors for their work.