Property	Value
Base model	sesame/csm-1b
Type	LoRA adapter
Language	Hindi (hi)
LoRA rank	32
LoRA alpha	64
Target modules	backbone_model.layers.\d+.self_attn.(q_proj
Dataset	Google FLEURS Hindi

Usage

Load as LoRA adapter (recommended)

python

from transformers import AutoProcessor, CsmForConditionalGeneration
from peft import PeftModel

base_model = CsmForConditionalGeneration.from_pretrained("sesame/csm-1b")
model = PeftModel.from_pretrained(base_model, "namankhurpia/csm-1b-hindi-lora")
model = model.merge_and_unload()
processor = AutoProcessor.from_pretrained("sesame/csm-1b")

model.eval()
model.to("cuda")  # or "mps" for Mac

# Generate Hindi speech
conversation = [
    {"role": "0", "content": [
        {"type": "text", "text": "नमस्ते, आज मौसम बहुत अच्छा है।"},
    ]},
    {"role": "1", "content": [
        {"type": "text", "text": "हाँ, बहुत सुहावना दिन है।"},
    ]},
]

inputs = processor.apply_chat_template(
    conversation, tokenize=True, return_dict=True,
)
inputs = {k: v.to(model.device) for k, v in inputs.items() if hasattr(v, "to")}

output = model.generate(**inputs, max_new_tokens=250, do_sample=False)
audio = processor.decode(output)

Architecture

Hindi text -> [CSM-1B + LoRA] -> Hindi audio
                    |
              Mimi codec (frozen)
              Backbone LM (LoRA on q/k/v/o_proj)
              Depth decoder (frozen)

Training Pipeline

Trained with the Hindi Speech-to-Speech Fine-Tuning pipeline.

License

This model is a fine-tuned adapter for sesame/csm-1b. See the base model's license for usage terms. Training data (FLEURS) is CC-BY-4.0.