gpt_sovits/feature_extractor/cnhubert.py

import torch

import logging

logging.getLogger("numba").setLevel(logging.WARNING)

from transformers import (
    Wav2Vec2FeatureExtractor,
    HubertModel,
)
import torch.nn as nn


class CNHubert(nn.Module):
    def __init__(self, cnhubert_base_path):
        super().__init__()
        self.model = HubertModel.from_pretrained(cnhubert_base_path)
        self.feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
            cnhubert_base_path
        )

    def forward(self, x):
        input_values = self.feature_extractor(
            x, return_tensors="pt", sampling_rate=16000
        ).input_values.to(x.device)
        feats = self.model(input_values)["last_hidden_state"]
        return feats


def get_content(hmodel, wav_16k_tensor):
    with torch.no_grad():
        feats = hmodel(wav_16k_tensor)
    return feats.transpose(1, 2)