Views
No views yet
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
Wav2Vec2ForCTC,
Wav2Vec2Processor,
AutoTokenizer,
AutoModelWithLMHead
)
import torch
import re
import sys
model_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
device = "cuda"
processor_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
chars_to_ignore_regex = r"[¥•＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､ 、〃〈〉《》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏﹑﹔·'℃°•·．﹑︰〈〉─《﹖﹣﹂﹁﹔！？｡。＂＃＄％＆＇（）＊＋，﹐－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏.．!\"#$%&()*+,\-.\:;<=>?@\[\]\\\/^_`{|}~]"
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2Processor.from_pretrained(processor_name)
tokenizer = AutoTokenizer.from_pretrained("ckiplab/gpt2-base-chinese")
gpt_model = AutoModelWithLMHead.from_pretrained("ckiplab/gpt2-base-chinese").to(device)
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
def load_file_to_data(file):
batch = {}
speech, _ = torchaudio.load(file)
batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
batch["sampling_rate"] = resampler.new_freq
return batch
def predict(data):
features = processor(data["speech"], sampling_rate=data["sampling_rate"], padding=True, return_tensors="pt")
input_values = features.input_values.to(device)
attention_mask = features.attention_mask.to(device)
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
decoded_results = []
for logit in logits:
pred_ids = torch.argmax(logit, dim=-1)
mask = pred_ids.ge(1).unsqueeze(-1).expand(logit.size())
vocab_size = logit.size()[-1]
voice_prob = torch.nn.functional.softmax((torch.masked_select(logit, mask).view(-1,vocab_size)),dim=-1)
gpt_input = torch.cat((torch.tensor([tokenizer.cls_token_id]).to(device),pred_ids[pred_ids>0]), 0)
gpt_prob = torch.nn.functional.softmax(gpt_model(gpt_input).logits, dim=-1)[:voice_prob.size()[0],:]
comb_pred_ids = torch.argmax(gpt_prob*voice_prob, dim=-1)
decoded_results.append(processor.decode(comb_pred_ids))
return decoded_resultspredict(load_file_to_data('voice file path'))!pip install editdistance
!pip install torchaudio
!pip install datasets transformers1import torchaudio
2from datasets import load_dataset, load_metric
3from transformers import (
4 Wav2Vec2ForCTC,
5 Wav2Vec2Processor,
6)
7import torch
8import re
9import sys
10from transformers import AutoTokenizer, AutoModelWithLMHead
11from datasets import Audio
12from math import log
13
14model_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
15device = "cuda"
16processor_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
17chars_to_ignore_regex = r"[¥•＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､ 、〃〈〉《》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏﹑﹔·'℃°•·．﹑︰〈〉─《﹖﹣﹂﹁﹔！？｡。＂＃＄％＆＇（）＊＋，﹐－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏.．!\"#$%&()*+,\-.\:;<=>?@\[\]\\\/^_`{|}~]"
18
19tokenizer = AutoTokenizer.from_pretrained("ckiplab/gpt2-base-chinese")
20lm_model = AutoModelWithLMHead.from_pretrained("ckiplab/gpt2-base-chinese").to(device)
21model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
22processor = Wav2Vec2Processor.from_pretrained(processor_name)
23
24ds = load_dataset("common_voice", 'zh-TW', split="test")
25ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
26def map_to_array(batch):
27 audio = batch["audio"]
28 batch["speech"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
29 batch["sampling_rate"] = audio["sampling_rate"]
30 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
31 return batch
32ds = ds.map(map_to_array)
33
34def map_to_pred(batch):
35 features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
36 input_values = features.input_values.to(device)
37 attention_mask = features.attention_mask.to(device)
38 with torch.no_grad():
39 logits = model(input_values, attention_mask=attention_mask).logits
40 pred_ids = torch.argmax(logits, dim=-1)
41 batch["predicted"] = processor.batch_decode(pred_ids)
42 batch["target"] = batch["sentence"]
43 return batch
44
45
46result = ds.map(map_to_pred, batched=True, batch_size=3, remove_columns=list(ds.features.keys()))
47
48def cer_cal(groundtruth, hypothesis):
49 err = 0
50 tot = 0
51 for p, t in zip(hypothesis, groundtruth):
52 err += float(ed.eval(p.lower(), t.lower()))
53 tot += len(t)
54 return err / tot
55print("CER: {:2f}".format(100 * cer_cal(result["target"],result["predicted"])))CER: 28.70.TIME: 04:08 min1import torchaudio
2from datasets import load_dataset, load_metric
3from transformers import (
4 Wav2Vec2ForCTC,
5 Wav2Vec2Processor,
6)
7import torch
8import re
9import sys
10from transformers import AutoTokenizer, AutoModelWithLMHead
11from datasets import Audio
12from math import log
13
14model_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
15device = "cuda"
16processor_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
17chars_to_ignore_regex = r"[¥•＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､ 、〃〈〉《》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏﹑﹔·'℃°•·．﹑︰〈〉─《﹖﹣﹂﹁﹔！？｡。＂＃＄％＆＇（）＊＋，﹐－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏.．!\"#$%&()*+,\-.\:;<=>?@\[\]\\\/^_`{|}~]"
18
19tokenizer = AutoTokenizer.from_pretrained("ckiplab/gpt2-base-chinese")
20lm_model = AutoModelWithLMHead.from_pretrained("ckiplab/gpt2-base-chinese").to(device)
21model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
22processor = Wav2Vec2Processor.from_pretrained(processor_name)
23
24ds = load_dataset("common_voice", 'zh-TW', split="test")
25ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
26def map_to_array(batch):
27 audio = batch["audio"]
28 batch["speech"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
29 batch["sampling_rate"] = audio["sampling_rate"]
30 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
31 return batch
32ds = ds.map(map_to_array)
33
34def map_to_pred(batch):
35 features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
36 input_values = features.input_values.to(device)
37 attention_mask = features.attention_mask.to(device)
38 with torch.no_grad():
39 logits = model(input_values, attention_mask=attention_mask).logits
40
41 decoded_results = []
42 for logit in logits:
43 pred_ids = torch.argmax(logit, dim=-1)
44 mask = pred_ids.ge(1).unsqueeze(-1).expand(logit.size())
45 vocab_size = logit.size()[-1]
46 voice_prob = torch.nn.functional.softmax((torch.masked_select(logit, mask).view(-1,vocab_size)),dim=-1)
47 lm_input = torch.cat((torch.tensor([tokenizer.cls_token_id]).to(device),pred_ids[pred_ids>0]), 0)
48 lm_prob = torch.nn.functional.softmax(lm_model(lm_input).logits, dim=-1)[:voice_prob.size()[0],:]
49 comb_pred_ids = torch.argmax(lm_prob*voice_prob, dim=-1)
50 decoded_results.append(processor.decode(comb_pred_ids))
51
52 batch["predicted"] = decoded_results
53 batch["target"] = batch["sentence"]
54 return batch
55
56
57result = ds.map(map_to_pred, batched=True, batch_size=3, remove_columns=list(ds.features.keys()))
58
59def cer_cal(groundtruth, hypothesis):
60 err = 0
61 tot = 0
62 for p, t in zip(hypothesis, groundtruth):
63 err += float(ed.eval(p.lower(), t.lower()))
64 tot += len(t)
65 return err / tot
66print("CER: {:2f}".format(100 * cer_cal(result["target"],result["predicted"])))CER 25.70.TIME: 06:04 min1import torchaudio
2from datasets import load_dataset, load_metric
3from transformers import (
4 Wav2Vec2ForCTC,
5 Wav2Vec2Processor,
6)
7import torch
8import re
9import sys
10from transformers import AutoTokenizer, AutoModelWithLMHead
11from datasets import Audio
12from math import log
13
14model_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
15device = "cuda"
16processor_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
17chars_to_ignore_regex = r"[¥•＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､ 、〃〈〉《》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏﹑﹔·'℃°•·．﹑︰〈〉─《﹖﹣﹂﹁﹔！？｡。＂＃＄％＆＇（）＊＋，﹐－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏.．!\"#$%&()*+,\-.\:;<=>?@\[\]\\\/^_`{|}~]"
18
19tokenizer = AutoTokenizer.from_pretrained("ckiplab/gpt2-base-chinese")
20lm_model = AutoModelWithLMHead.from_pretrained("ckiplab/gpt2-base-chinese").to(device)
21model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
22processor = Wav2Vec2Processor.from_pretrained(processor_name)
23
24ds = load_dataset("common_voice", 'zh-TW', split="test")
25ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
26def map_to_array(batch):
27 audio = batch["audio"]
28 batch["speech"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
29 batch["sampling_rate"] = audio["sampling_rate"]
30 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
31 return batch
32ds = ds.map(map_to_array)
33
34def map_to_pred(batch):
35 features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
36 input_values = features.input_values.to(device)
37 attention_mask = features.attention_mask.to(device)
38 with torch.no_grad():
39 logits = model(input_values, attention_mask=attention_mask).logits
40
41 decoded_results = []
42 for logit in logits:
43 sequences = [[[], 1.0]]
44 pred_ids = torch.argmax(logit, dim=-1)
45 mask = pred_ids.ge(1).unsqueeze(-1).expand(logit.size())
46 vocab_size = logit.size()[-1]
47 voice_prob = torch.nn.functional.softmax((torch.masked_select(logit, mask).view(-1,vocab_size)),dim=-1)
48 while True:
49 all_candidates = list()
50 exceed = False
51 for seq in sequences:
52 tokens, score = seq
53 gpt_input = torch.tensor([tokenizer.cls_token_id]+tokens).to(device)
54 gpt_prob = torch.nn.functional.softmax(lm_model(gpt_input).logits, dim=-1)[:len(gpt_input),:]
55 if len(gpt_input) >= len(voice_prob):
56 exceed = True
57 comb_pred_ids = gpt_prob*voice_prob[:len(gpt_input)]
58 v,i = torch.topk(comb_pred_ids,50,dim=-1)
59 for tok_id,tok_prob in zip(i.tolist()[-1],v.tolist()[-1]):
60 candidate = [tokens + [tok_id], score + -log(tok_prob)]
61 all_candidates.append(candidate)
62 ordered = sorted(all_candidates, key=lambda tup: tup[1])
63 sequences = ordered[:10]
64 if exceed:
65 break
66 decoded_results.append(processor.decode(sequences[0][0]))
67
68 batch["predicted"] = decoded_results
69 batch["target"] = batch["sentence"]
70 return batch
71
72
73result = ds.map(map_to_pred, batched=True, batch_size=3, remove_columns=list(ds.features.keys()))
74
75def cer_cal(groundtruth, hypothesis):
76 err = 0
77 tot = 0
78 for p, t in zip(hypothesis, groundtruth):
79 err += float(ed.eval(p.lower(), t.lower()))
80 tot += len(t)
81 return err / tot
82print("CER: {:2f}".format(100 * cer_cal(result["target"],result["predicted"])))CER 18.36.1import torchaudio
2from datasets import load_dataset, load_metric
3from transformers import (
4 Wav2Vec2ForCTC,
5 Wav2Vec2Processor,
6)
7import torch
8import re
9import sys
10from transformers import AutoTokenizer, AutoModelForMaskedLM
11
12model_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
13device = "cuda"
14processor_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
15chars_to_ignore_regex = r"[¥•＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､ 、〃〈〉《》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏﹑﹔·'℃°•·．﹑︰〈〉─《﹖﹣﹂﹁﹔！？｡。＂＃＄％＆＇（）＊＋，﹐－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏.．!\"#$%&()*+,\-.\:;<=>?@\[\]\\\/^_`{|}~]"
16
17tokenizer = AutoTokenizer.from_pretrained("bert-base-chinese")
18lm_model = AutoModelForMaskedLM.from_pretrained("bert-base-chinese").to(device)
19model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
20processor = Wav2Vec2Processor.from_pretrained(processor_name)
21
22ds = load_dataset("common_voice", 'zh-TW', data_dir="./cv-corpus-6.1-2020-12-11", split="test")
23
24resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
25
26def map_to_array(batch):
27 speech, _ = torchaudio.load(batch["path"])
28 batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
29 batch["sampling_rate"] = resampler.new_freq
30 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
31 return batch
32
33ds = ds.map(map_to_array)
34
35def map_to_pred(batch):
36 features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
37 input_values = features.input_values.to(device)
38 attention_mask = features.attention_mask.to(device)
39 with torch.no_grad():
40 logits = model(input_values, attention_mask=attention_mask).logits
41
42 decoded_results = []
43 for logit in logits:
44 pred_ids = torch.argmax(logit, dim=-1)
45 mask = ~pred_ids.eq(tokenizer.pad_token_id).unsqueeze(-1).expand(logit.size())
46 vocab_size = logit.size()[-1]
47 voice_prob = torch.nn.functional.softmax((torch.masked_select(logit, mask).view(-1,vocab_size)),dim=-1)
48 lm_input = torch.masked_select(pred_ids, ~pred_ids.eq(tokenizer.pad_token_id)).unsqueeze(0)
49 mask_lm_prob = voice_prob.clone()
50 for i in range(lm_input.shape[-1]):
51 masked_lm_input = lm_input.clone()
52 masked_lm_input[0][i] = torch.tensor(tokenizer.mask_token_id).to('cuda')
53 lm_prob = torch.nn.functional.softmax(lm_model(masked_lm_input).logits, dim=-1).squeeze(0)
54 mask_lm_prob[i] = lm_prob[i]
55 comb_pred_ids = torch.argmax(mask_lm_prob*voice_prob, dim=-1)
56 decoded_results.append(processor.decode(comb_pred_ids))
57
58 batch["predicted"] = decoded_results
59 batch["target"] = batch["sentence"]
60 return batch
61
62
63result = ds.map(map_to_pred, batched=True, batch_size=1, remove_columns=list(ds.features.keys()))
64
65def cer_cal(groundtruth, hypothesis):
66 err = 0
67 tot = 0
68 for p, t in zip(hypothesis, groundtruth):
69 err += float(ed.eval(p.lower(), t.lower()))
70 tot += len(t)
71 return err / tot
72print("CER: {:2f}".format(100 * cer_cal(result["target"],result["predicted"])))CER 25.57.TIME: 09:49 min!git clone https://github.com/voidful/pytorch-tta.git
!mv ./pytorch-tta/tta ./tta
!wget https://github.com/voidful/pytorch-tta/releases/download/wiki_zh/wiki_zh.pt1import torchaudio
2from datasets import load_dataset, load_metric
3from transformers import (
4 Wav2Vec2ForCTC,
5 Wav2Vec2Processor,
6)
7import torch
8import re
9import sys
10from tta.modeling_tta import TTALMModel
11from transformers import AutoTokenizer
12import torch
13
14
15
16model_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
17device = "cuda"
18processor_name = "voidful/wav2vec2-large-xlsr-53-tw-gpt"
19chars_to_ignore_regex = r"[¥•＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､ 、〃〈〉《》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏﹑﹔·'℃°•·．﹑︰〈〉─《﹖﹣﹂﹁﹔！？｡。＂＃＄％＆＇（）＊＋，﹐－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘’‛“”„‟…‧﹏.．!\"#$%&()*+,\-.\:;<=>?@\[\]\\\/^_`{|}~]"
20
21tokenizer = AutoTokenizer.from_pretrained("bert-base-chinese")
22lm_model = TTALMModel("bert-base-chinese")
23tokenizer = AutoTokenizer.from_pretrained("bert-base-chinese")
24lm_model.load_state_dict(torch.load("./wiki_zh.pt",map_location=torch.device('cuda')))
25lm_model.to('cuda')
26lm_model.eval()
27model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
28processor = Wav2Vec2Processor.from_pretrained(processor_name)
29
30ds = load_dataset("common_voice", 'zh-TW', data_dir="./cv-corpus-6.1-2020-12-11", split="test")
31
32resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
33
34def map_to_array(batch):
35 speech, _ = torchaudio.load(batch["path"])
36 batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
37 batch["sampling_rate"] = resampler.new_freq
38 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
39 return batch
40
41ds = ds.map(map_to_array)
42
43def map_to_pred(batch):
44 features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
45 input_values = features.input_values.to(device)
46 attention_mask = features.attention_mask.to(device)
47 with torch.no_grad():
48 logits = model(input_values, attention_mask=attention_mask).logits
49
50 decoded_results = []
51 for logit in logits:
52 pred_ids = torch.argmax(logit, dim=-1)
53 mask = ~pred_ids.eq(tokenizer.pad_token_id).unsqueeze(-1).expand(logit.size())
54 vocab_size = logit.size()[-1]
55 voice_prob = torch.nn.functional.softmax((torch.masked_select(logit, mask).view(-1,vocab_size)),dim=-1)
56 lm_input = torch.masked_select(pred_ids, ~pred_ids.eq(tokenizer.pad_token_id)).unsqueeze(0)
57 lm_prob = torch.nn.functional.softmax(lm_model.forward(lm_input)[0], dim=-1).squeeze(0)
58 comb_pred_ids = torch.argmax(lm_prob*voice_prob, dim=-1)
59 decoded_results.append(processor.decode(comb_pred_ids))
60
61 batch["predicted"] = decoded_results
62 batch["target"] = batch["sentence"]
63 return batch
64
65
66result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
67
68def cer_cal(groundtruth, hypothesis):
69 err = 0
70 tot = 0
71 for p, t in zip(hypothesis, groundtruth):
72 err += float(ed.eval(p.lower(), t.lower()))
73 tot += len(t)
74 return err / tot
75print("CER: {:2f}".format(100 * cer_cal(result["target"],result["predicted"])))CER: 25.77.TIME: 06:01 min