File size: 4,135 Bytes
8284447
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
52cd932
8284447
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
52cd932
8284447
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
---
language: sah
datasets:
- common_voice
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: Sakha XLSR Wav2Vec2 Large 53 by Anton Lozhkov
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice sah
      type: common_voice
      args: sah
    metrics:
       - name: Test WER
         type: wer
         value: 32.23
---

# Wav2Vec2-Large-XLSR-53-Sakha

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.
When using this model, make sure that your speech input is sampled at 16kHz.

## Usage

The model can be used directly (without a language model) as follows:

```python
import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

test_dataset = load_dataset("common_voice", "sah", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])
```


## Evaluation

The model can be evaluated as follows on the Sakha test data of Common Voice.

```python
import torch
import torchaudio
import urllib.request
import tarfile
import pandas as pd
from tqdm.auto import tqdm
from datasets import load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# Download the raw data instead of using HF datasets to save disk space 
data_url = "https://voice-prod-bundler-ee1969a6ce8178826482b88e843c335139bd3fb4.s3.amazonaws.com/cv-corpus-6.1-2020-12-11/sah.tar.gz"
filestream = urllib.request.urlopen(data_url)
data_file = tarfile.open(fileobj=filestream, mode="r|gz")
data_file.extractall()

wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-sakha")
model.to("cuda")

cv_test = pd.read_csv("cv-corpus-6.1-2020-12-11/sah/test.tsv", sep='\t')
clips_path = "cv-corpus-6.1-2020-12-11/sah/clips/"

def clean_sentence(sent):
    sent = sent.lower()
    # replace non-alpha characters with space
    sent = "".join(ch if ch.isalpha() else " " for ch in sent)
    # remove repeated spaces
    sent = " ".join(sent.split())
    return sent

targets = []
preds = []

for i, row in tqdm(cv_test.iterrows(), total=cv_test.shape[0]):
    row["sentence"] = clean_sentence(row["sentence"])
    speech_array, sampling_rate = torchaudio.load(clips_path + row["path"])
    resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
    row["speech"] = resampler(speech_array).squeeze().numpy()

    inputs = processor(row["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits

    pred_ids = torch.argmax(logits, dim=-1)

    targets.append(row["sentence"])
    preds.append(processor.batch_decode(pred_ids)[0])

print("WER: {:2f}".format(100 * wer.compute(predictions=preds, references=targets)))
```

**Test Result**: 32.23 %  


## Training

The Common Voice `train` and `validation` datasets were used for training.