from transformers import Wav2Vec2Processor, HubertModel import soundfile as sf import numpy as np import torch print("Loading the Wav2Vec2 Processor...") wav2vec2_processor = Wav2Vec2Processor.from_pretrained("facebook/hubert-large-ls960-ft") print("Loading the HuBERT Model...") hubert_model = HubertModel.from_pretrained("facebook/hubert-large-ls960-ft") def get_hubert_from_16k_wav(wav_16k_name): speech_16k, _ = sf.read(wav_16k_name) hubert = get_hubert_from_16k_speech(speech_16k) return hubert @torch.no_grad() def get_hubert_from_16k_speech(speech, device="cuda:0"): global hubert_model hubert_model = hubert_model.to(device) if speech.ndim ==2: speech = speech[:, 0] # [T, 2] ==> [T,] input_values_all = wav2vec2_processor(speech, return_tensors="pt", sampling_rate=16000).input_values # [1, T] input_values_all = input_values_all.to(device) # For long audio sequence, due to the memory limitation, we cannot process them in one run # HuBERT process the wav with a CNN of stride [5,2,2,2,2,2], making a stride of 320 # Besides, the kernel is [10,3,3,3,3,2,2], making 400 a fundamental unit to get 1 time step. # So the CNN is euqal to a big Conv1D with kernel k=400 and stride s=320 # We have the equation to calculate out time step: T = floor((t-k)/s) # To prevent overlap, we set each clip length of (K+S*(N-1)), where N is the expected length T of this clip # The start point of next clip should roll back with a length of (kernel-stride) so it is stride * N kernel = 400 stride = 320 clip_length = stride * 1000 num_iter = input_values_all.shape[1] // clip_length expected_T = (input_values_all.shape[1] - (kernel-stride)) // stride res_lst = [] for i in range(num_iter): if i == 0: start_idx = 0 end_idx = clip_length - stride + kernel else: start_idx = clip_length * i end_idx = start_idx + (clip_length - stride + kernel) input_values = input_values_all[:, start_idx: end_idx] hidden_states = hubert_model.forward(input_values).last_hidden_state # [B=1, T=pts//320, hid=1024] res_lst.append(hidden_states[0]) if num_iter > 0: input_values = input_values_all[:, clip_length * num_iter:] else: input_values = input_values_all # if input_values.shape[1] != 0: if input_values.shape[1] >= kernel: # if the last batch is shorter than kernel_size, skip it hidden_states = hubert_model(input_values).last_hidden_state # [B=1, T=pts//320, hid=1024] res_lst.append(hidden_states[0]) ret = torch.cat(res_lst, dim=0).cpu() # [T, 1024] # assert ret.shape[0] == expected_T assert abs(ret.shape[0] - expected_T) <= 1 if ret.shape[0] < expected_T: ret = torch.nn.functional.pad(ret, (0,0,0,expected_T-ret.shape[0])) else: ret = ret[:expected_T] return ret def make_even_first_dim(tensor): size = list(tensor.size()) if size[0] % 2 == 1: size[0] -= 1 return tensor[:size[0]] return tensor import soundfile as sf import numpy as np import torch from argparse import ArgumentParser import librosa parser = ArgumentParser() parser.add_argument('--wav', type=str, help='') args = parser.parse_args() wav_name = args.wav speech, sr = sf.read(wav_name) speech_16k = librosa.resample(speech, orig_sr=sr, target_sr=16000) print("SR: {} to {}".format(sr, 16000)) # print(speech.shape, speech_16k.shape) hubert_hidden = get_hubert_from_16k_speech(speech_16k) hubert_hidden = make_even_first_dim(hubert_hidden).reshape(-1, 2, 1024) np.save(wav_name.replace('.wav', '_hu.npy'), hubert_hidden.detach().numpy()) print(hubert_hidden.detach().numpy().shape)