csm-podcast-finetuned-en

CSM Podcast Finetuned (English)

This model is a finetuned version of sesame/csm-1b (Conversational Speech Model) trained on the Vyvo/en-podcast-1 dataset.

Training Details

• Base Model: sesame/csm-1b
• Dataset: Vyvo/en-podcast-1 (68,891 samples)
• Training Method: LoRA finetuning
• Epochs: 4
• Final Loss: 6.77 (from initial 7.56, ~11% improvement)
• LoRA Config: rank=32, alpha=32

Usage

from generator import Generator
from models import Model
from safetensors.torch import load_file
import torch

# Load model
model = Model.from_pretrained("Codyfederer/csm-podcast-finetuned-en")
model.to(device="cuda", dtype=torch.bfloat16)

# Create generator
generator = Generator(model)

# Generate audio
audio = generator.generate(
    text="Hello, this is a test of the finetuned model.",
    speaker=0,
    context=[],
    max_audio_length_ms=10000,
)

Model Files

• model.safetensors - The finetuned model weights (merged LoRA)
• config.json - Model configuration

License

Apache 2.0