for i, (x, y, x_lens, y_lens) in tqdm(enumerate(dataloader)):
# x = x.to("cuda")
x = x.to(device)
outs, out_lens = model(x, x_lens)
outs = F.softmax(outs, 1)
outs = outs.transpose(1, 2)
ys = []
offset = 0
for y_len in y_lens:
ys.append(y[offset : offset + y_len])
offset += y_len
out_strings, out_offsets = decoder.decode(outs, out_lens)
y_strings = decoder.convert_to_strings(ys)
for pred, truth in zip(out_strings, y_strings):
trans, ref = pred[0], truth[0]
cer += decoder.cer(trans, ref) / float(len(ref))
cer /= len(dataloader.dataset)
model.train()
return cer
if __name__ == "__main__":
with open("./labels.json") as f:
vocabulary = json.load(f)
vocabulary = "".join(vocabulary)
model = GatedConv(vocabulary)
if(not os.path.exists(save_path)):
os.mkdir(save_path)
model.to(device)
train(model)
import torch
import feature
from models.conv import GatedConv
import torch.nn.functional as F
from ctcdecode import CTCBeamDecoder
from config import lm_path, pretrained_model_path
alpha = 0.8
beta = 0.3
cutoff_top_n = 40
cutoff_prob = 1.0
beam_width = 32
num_processes = 4
blank_index = 0
model = GatedConv.load(pretrained_model_path)
model.eval()
decoder = CTCBeamDecoder(
model.vocabulary,
lm_path,
alpha,
beta,
cutoff_top_n,
cutoff_prob,
beam_width,
num_processes,
blank_index,
)
import _init_path
import platform
from models.conv import GatedConv
use_lm = True
if use_lm:
import beamdecode
system_type = platform.system()
if (system_type == 'Windows'):
model = GatedConv.load("AboutDL\\语音识别MASR\\pretrained\\gated-conv.pth")
#import scipy
#_,receipt_data = scipy.io.wavfile.read("E:\\打开欢呼比.wav")
#text = model.predict(receipt_data)事实证明效果相同
text = model.predict("E:\\打开欢呼比.wav")
elif (system_type == 'Linux'):
model = GatedConv.load('AboutDL/语音识别MASR/pretrained/gated-conv.pth')
text = model.predict(
"/media/yangjinming/DATA/Dataset/PrimeWords/d/d2/d25104a2-6be0-4950-9ec0-42e8e1303492.wav"
)
print("识别结果:", text)
print("decoding")
with torch.no_grad():
for i, (x, y, x_lens, y_lens) in tqdm(enumerate(dataloader)):
x = x.to(device)
outs, out_lens = model(x, x_lens)
outs = F.softmax(outs, 1)
outs = outs.transpose(1, 2)
ys = []
offset = 0
for y_len in y_lens:
ys.append(y[offset:offset + y_len])
offset += y_len
out_strings, out_offsets = decoder.decode(outs, out_lens)
y_strings = decoder.convert_to_strings(ys)
for pred, truth in zip(out_strings, y_strings):
trans, ref = pred[0], truth[0]
cer += decoder.cer(trans, ref) / float(len(ref))
cer /= len(dataloader.dataset)
model.train()
return cer
if __name__ == "__main__":
model = GatedConv(
json.load(open("../data_aishell/labels.json", encoding='utf-8')))
epoch = 40
model.load_state_dict(torch.load("pretrained/model_{}.pth".format(epoch)))
print("reload model: pretrained/model_{}.pth".format(epoch))
model.to(device)
train(model, start_epoch=epoch)
import _init_path
from models.conv import GatedConv
import json
# model = GatedConv.load("pretrained/gated-conv.pth")
# model.to_train()
# model.fit("train.manifest", "train.manifest")
with open("data_aishell/labels.json") as f:
vocabulary = json.load(f)
vocabulary = "".join(vocabulary)
model = GatedConv(vocabulary)
model.to_train()
model.fit("/home/dolan/Desktop/masr/data_aishell/train.index",
"/home/dolan/Desktop/masr/data_aishell/dev.index",
"/home/dolan/Desktop/masr/data_aishell/labels.json", 10)
import _init_path
from models.conv import GatedConv
import pre_transform_2 as pt2
import enhance_speach as es
# import beamdecode
model = GatedConv.load("pretrained/gated-conv.pth")
# model = GatedConv.load("pretrained/model_110.pth")
# with open("train_index","w") as f:
# for filename in os.listdir('data/'):
# # print(filename)
# s1 = filename[:8]
# s2 = ""
# for x in s1:
# s2 += trans[x]
# s1 = "data/" + s1 +".wav"
# f.write(s1+","+s2+"\n")
# text = model.predict("test.wav")
# text = model.predict("12345_man.wav")
# text = model.predict("678910_man.wav")
# text = model.predict("862409_in.wav")
# text = model.predict("20164239_kuai.wav")
# text = model.predict("20164762_kuai.wav")
# text = model.predict("20164762_man.wav")
# text = model.predict("data/20166565.wav")
# text = model.predict("20164786.wav")
input_file_src = "record.wav"
# output_file_src = "record_out.wav"
# es.denoise(input_file_src,output_file_src)
# import _init_path
from models.conv import GatedConv
# model = GatedConv.load("pretrained/gated-conv.pth")
# model = GatedConv.load("pretrained/model_62.pth")
model = GatedConv.load("pretrained2/model_81.pth")
text = model.predict("./sample_audio/8_16.wav")
print("")
print("识别结果:")
print(text)
import _init_path
from models.conv import GatedConv
model = GatedConv.load("语音识别MASR/pretrained/gated-conv.pth")
model.to_train()
model.fit("data/train.index", "data/dev.index", train_batch_size=2)
__mtime__ = '20210318'
import os
from models.conv import GatedConv
from config import pretrained_model_path
model = GatedConv.load(os.path.join('..', pretrained_model_path))
text = model.predict("../data_aishell/BAC009S0765W0130.wav")
print("")
print("识别结果:")
print(text)
cer = 0
print("decoding")
with torch.no_grad():
for i, (x, y, x_lens, y_lens) in tqdm(enumerate(dataloader)):
x = x.to("cuda")
outs, out_lens = model(x, x_lens)
outs = F.softmax(outs, 1)
outs = outs.transpose(1, 2)
ys = []
offset = 0
for y_len in y_lens:
ys.append(y[offset : offset + y_len])
offset += y_len
out_strings, out_offsets = decoder.decode(outs, out_lens)
y_strings = decoder.convert_to_strings(ys)
for pred, truth in zip(out_strings, y_strings):
trans, ref = pred[0], truth[0]
cer += decoder.cer(trans, ref) / float(len(ref))
cer /= len(dataloader.dataset)
model.train()
return cer
if __name__ == "__main__":
with open("data_aishell/labels.json") as f:
vocabulary = json.load(f)
vocabulary = "".join(vocabulary)
model = GatedConv(vocabulary)
model.to("cuda")
train(model)
import torch
import feature
from models.conv import GatedConv
import torch.nn.functional as F
from ctcdecode import CTCBeamDecoder
alpha = 0.8
beta = 0.3
lm_path = "/home/db/bing/yuyingshibie/masr/lm/zh_giga.no_cna_cmn.prune01244.klm"
cutoff_top_n = 40
cutoff_prob = 1.0
beam_width = 32
num_processes = 4
blank_index = 0
model = GatedConv.load(
"/home/db/bing/yuyingshibie/masr/pretrained/gated-conv.pth")
model.eval()
decoder = CTCBeamDecoder(
model.vocabulary,
lm_path,
alpha,
beta,
cutoff_top_n,
cutoff_prob,
beam_width,
num_processes,
blank_index,
)
import _init_path
from models.conv import GatedConv
# model = GatedConv.load("pretrained/gated-conv.pth")
model = GatedConv.load("pretrained/model_3.pth")
text = model.predict("./sample_audio/test.wav")
print("")
print("识别结果:")
print(text)
parser.add_argument('--lm-alpha-from', default=0.0, type=float, help='Language model weight start tuning')
parser.add_argument('--lm-alpha-to', default=3.0, type=float, help='Language model weight end tuning')
parser.add_argument('--lm-beta-from', default=0.0, type=float,
help='Language model word bonus (all words) start tuning')
parser.add_argument('--lm-beta-to', default=0.5, type=float,
help='Language model word bonus (all words) end tuning')
parser.add_argument('--lm-num-alphas', default=45, type=float, help='Number of alpha candidates for tuning')
parser.add_argument('--lm-num-betas', default=5, type=float, help='Number of beta candidates for tuning')
parser = add_decoder_args(parser)
args = parser.parse_args()
if args.lm_path is None:
print("error: LM must be provided for tuning")
sys.exit(1)
model = GatedConv.load(args.model_path)
saved_output = np.load(args.saved_output, allow_pickle=True)
def init(beam_width, blank_index, lm_path):
global decoder, ae_decoder
decoder = BeamCTCDecoder(model.vocabulary, lm_path=lm_path, beam_width=beam_width, num_processes=args.lm_workers,
blank_index=blank_index)
ae_decoder = GreedyDecoder(model.vocabulary)
def decode_dataset(params):
lm_alpha, lm_beta = params
global decoder
decoder._decoder.reset_params(lm_alpha, lm_beta)