facebook

/

mms-tts-som

like 10

Follow

AI at Meta 13.6k

Hey @adityaedy01 - this is a text-to-speech model. Therefore, it uses a tokenizer to pre-process the text inputs to token ids. It doesn't need a preprocessor_config.json, which is only used to define speech feature extractors (speech -> input features). That means in practice, you can use the tokenizer directly to as follows, without a need for a feature extractor:

from transformers import VitsModel, AutoTokenizer
import torch

model = VitsModel.from_pretrained("facebook/mms-tts-som")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-som")

text = "some example text in the Somali language"

# pre-processing of input text to token ids
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    output = model(**inputs).waveform

Hey @adityaedy01 - this is a text-to-speech model. Therefore, it uses a tokenizer to pre-process the text inputs to token ids. It doesn't need a preprocessor_config.json, which is only used to define speech feature extractors (speech -> input features). That means in practice, you can use the tokenizer directly to as follows, without a need for a feature extractor:
<
You're talking complete nonsense. He's talking about training, not just generating audio. Those are two completely different things. The file is missing because this is not a checkpoint @adityaedy01 . You need a checkpoint for further training. This model is clean and only for audio generation. You need this from the link below, albeit with a correction to get the checkpoint where this file is. https://github.com/ylacombe/finetune-hf-vits