Finetune on custom voice using the snorbyte/snorTTS-Indic-v0

#3
by Shekharmeena - opened

so I have single speaker High quality 30 hours of data, I have doubt can I fin tune your finetuned model on top of my voice dataset, as the script provided by u is from the base model of orpheus , I want to use your finetuned snorbyte/snorTTS-Indic-v0 model to further finetune on my custom voice. what parameters do I need to change, what should I do?

Snorbyte org

Hi, you have to format the dataset in SNAC encoding. Please refer to https://huggingface.co/datasets/snorbyte/indic-tts-sample-snac-encoded to see the format. Only utterance, user and language and SNAC codes are mandatory.

To finetune on your dataset just make the following changes. I have marked the changes below.

# Set up constants and configurations.
STAGE = 2 # <---------- changed
HUGGINGFACE_USERNAME = ""  # ! Fill.

if STAGE == 1:
    # No changes required ...
else:
    BASE_MODEL = "snorbyte/snorTTS-Indic-v0" # <---------- changed
    TARGET_MODULES = [
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
        "up_proj",
        "down_proj",
        "gate_proj",
    ]
    TRAIN_CSV_PATH =  ""  # ! Fill.
    VALID_CSV_PATH = ""  # ! Fill.
    LR = 2e-4
    EPOCHS = 2
    MODEL_NAME = "snorTTS-indicv0-custom" # <---------- changed
SaudxInu changed discussion status to closed

Sign up or log in to comment