| import time |
|
|
| import librosa |
| import torch |
| import torch.nn.functional as F |
| import soundfile as sf |
| import logging |
|
|
| logging.getLogger("numba").setLevel(logging.WARNING) |
|
|
| from transformers import ( |
| Wav2Vec2FeatureExtractor, |
| HubertModel, |
| ) |
| from torchaudio.transforms import Resample |
|
|
| import utils |
| import torch.nn as nn |
|
|
| class CNHubert(nn.Module): |
| def __init__(self, input_sample_rate=32000): |
| super().__init__() |
| self.model = HubertModel.from_pretrained("TencentGameMate/chinese-hubert-base") |
| self.feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("TencentGameMate/chinese-hubert-base") |
| if input_sample_rate != 16000: |
| self.resample = Resample(orig_freq=input_sample_rate, new_freq=16000) |
| else: |
| self.resample = None |
| self.input_sample_rate = input_sample_rate |
| self.log_flag = False |
| def forward(self, x): |
| if self.resample is not None: |
| if not self.log_flag: |
| print(f"Resampling from {x.shape[-1]} to 16000") |
| self.log_flag = True |
| x = self.resample(x) |
| input_values = self.feature_extractor(x, return_tensors="pt", sampling_rate=16000).input_values.to(x.device).to(x.dtype) |
| feats = self.model(input_values)["last_hidden_state"] |
| return feats |
|
|
|
|
|
|
| def get_model(input_sample_rate=32000): |
| model = CNHubert(input_sample_rate=input_sample_rate) |
| model.eval() |
| return model |
|
|
| def get_content(hmodel, src_path): |
| input_sample_rate = hmodel.input_sample_rate |
| wav, _ = librosa.load(src_path, sr=input_sample_rate) |
| device = hmodel.parameters().__next__().device |
| dtype = hmodel.parameters().__next__().dtype |
| wav_16k_tensor = torch.from_numpy(wav).to(device).to(dtype) |
| with torch.no_grad(): |
| feats = hmodel(wav_16k_tensor) |
| return feats.transpose(1,2) |
|
|
|
|
| if __name__ == '__main__': |
| model = get_model() |
| src_path = "/Users/Shared/原音频2.wav" |
| wav_16k_tensor = utils.load_wav_to_torch_and_resample(src_path, 16000) |
| model = model |
| wav_16k_tensor = wav_16k_tensor |
| feats = get_content(model,wav_16k_tensor) |
| print(feats.shape) |
|
|
|
|
