speech-test commited on
Commit
8284447
1 Parent(s): 7fe3e81

Add model card

Browse files
Files changed (1) hide show
  1. README.md +138 -1
README.md CHANGED
@@ -1 +1,138 @@
1
- TODO
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: sah
3
+ datasets:
4
+ - common_voice
5
+ metrics:
6
+ - wer
7
+ tags:
8
+ - audio
9
+ - automatic-speech-recognition
10
+ - speech
11
+ - xlsr-fine-tuning-week
12
+ license: apache-2.0
13
+ model-index:
14
+ - name: Sakha XLSR Wav2Vec2 Large 53 by Anton Lozhkov
15
+ results:
16
+ - task:
17
+ name: Speech Recognition
18
+ type: automatic-speech-recognition
19
+ dataset:
20
+ name: Common Voice sah
21
+ type: common_voice
22
+ args: sah
23
+ metrics:
24
+ - name: Test WER
25
+ type: wer
26
+ value: 32.23
27
+ ---
28
+
29
+ # Wav2Vec2-Large-XLSR-53-Sakha
30
+
31
+ Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Sakha using the [Common Voice](https://huggingface.co/datasets/common_voice) dataset.
32
+ When using this model, make sure that your speech input is sampled at 16kHz.
33
+
34
+ ## Usage
35
+
36
+ The model can be used directly (without a language model) as follows:
37
+
38
+ ```python
39
+ import torch
40
+ import torchaudio
41
+ from datasets import load_dataset
42
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
43
+
44
+ test_dataset = load_dataset("common_voice", "sah", split="test[:2%]")
45
+
46
+ processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
47
+ model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
48
+
49
+ resampler = torchaudio.transforms.Resample(48_000, 16_000)
50
+
51
+ # Preprocessing the datasets.
52
+ # We need to read the audio files as arrays
53
+ def speech_file_to_array_fn(batch):
54
+ speech_array, sampling_rate = torchaudio.load(batch["path"])
55
+ batch["speech"] = resampler(speech_array).squeeze().numpy()
56
+ return batch
57
+
58
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
59
+ inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
60
+
61
+ with torch.no_grad():
62
+ logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
63
+
64
+ predicted_ids = torch.argmax(logits, dim=-1)
65
+
66
+ print("Prediction:", processor.batch_decode(predicted_ids))
67
+ print("Reference:", test_dataset["sentence"][:2])
68
+ ```
69
+
70
+
71
+ ## Evaluation
72
+
73
+ The model can be evaluated as follows on the Sakha test data of Common Voice.
74
+
75
+ ```python
76
+ import torch
77
+ import torchaudio
78
+ import urllib.request
79
+ import tarfile
80
+ import pandas as pd
81
+ from tqdm.auto import tqdm
82
+ from datasets import load_metric
83
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
84
+
85
+ # Download the raw data instead of using HF datasets to save space
86
+ data_url = "https://voice-prod-bundler-ee1969a6ce8178826482b88e843c335139bd3fb4.s3.amazonaws.com/cv-corpus-6.1-2020-12-11/sah.tar.gz"
87
+ filestream = urllib.request.urlopen(data_url)
88
+ data_file = tarfile.open(fileobj=filestream, mode="r|gz")
89
+ data_file.extractall()
90
+
91
+ wer = load_metric("wer")
92
+
93
+ processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
94
+ model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
95
+ model.to("cuda")
96
+
97
+ cv_test = pd.read_csv("cv-corpus-6.1-2020-12-11/sah/test.tsv", sep='\t')
98
+ clips_path = "cv-corpus-6.1-2020-12-11/sah/clips/"
99
+
100
+ def clean_sentence(sent):
101
+ sent = sent.lower()
102
+ # replace non-alpha characters with space
103
+ sent = "".join(ch if ch.isalpha() else " " for ch in sent)
104
+ # remove repeated spaces
105
+ sent = " ".join(sent.split())
106
+ return sent
107
+
108
+ resampler = torchaudio.transforms.Resample(48_000, 16_000)
109
+
110
+ targets = []
111
+ preds = []
112
+
113
+ for i, row in tqdm(cv_test.iterrows(), total=cv_test.shape[0]):
114
+ row["sentence"] = clean_sentence(row["sentence"])
115
+ speech_array, sampling_rate = torchaudio.load(clips_path + row["path"])
116
+ row["speech"] = resampler(speech_array).squeeze().numpy()
117
+
118
+ inputs = processor(row["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
119
+
120
+ with torch.no_grad():
121
+ logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
122
+
123
+ pred_ids = torch.argmax(logits, dim=-1)
124
+
125
+ targets.append(row["sentence"])
126
+ preds.append(processor.batch_decode(pred_ids)[0])
127
+
128
+ print("WER: {:2f}".format(100 * wer.compute(predictions=preds, references=targets)))
129
+ ```
130
+
131
+ **Test Result**: 32.23 %
132
+
133
+
134
+ ## Training
135
+
136
+ The Common Voice `train` and `validation` datasets were used for training.
137
+
138
+ The script used for training can be found [here](github.com)