Aaryan39 / hinglish-tts-3b-ft

Hinglish TTS 3B Model

This is a fine-tuned version of canopylabs/3b-hi-pretrain-research_release specialized for Hinglish (Hindi-English mixed) text-to-speech generation.

Model Details

• Base Model : canopylabs/3b-hi-pretrain-research_release
• Fine-tuning Method : LoRA with Unsloth (merged)
• Languages : Hindi, English, Hinglish
• Task : Text-to-Speech via audio token generation
• Model Size : ~3B parameters

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
model_name = "Aaryan39/hinglish-tts-3b-ft"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Generate text
prompt = "Hello doston, main aapka dost hun"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=1200)

Fine-tuning Details

• LoRA Rank : 64
• LoRA Alpha : 64
• Target Modules : q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Framework : Unsloth

Audio Generation

This model generates audio tokens that need to be decoded using a SNAC (Scalable Neural Audio Codec) model:

from snac import SNAC

# Load SNAC decoder
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")

# Process generated tokens to audio codes and decode
# (See full implementation in the original training code)

Limitations

• Requires SNAC model for audio generation
• Optimized for Hinglish content
• May not perform well on pure English or pure Hindi in some cases

Citation

If you use this model, please cite the original base model:

@misc{canopylabs-3b-hi,
  title={3B Hindi Pretrained Model},
  author={Canopy Labs},
  year={2024},
  url={https://huggingface.co/canopylabs/3b-hi-pretrain-research_release}
}