File size: 4,717 Bytes
f8903fb
 
 
0c63c7e
f8903fb
 
 
 
 
 
 
 
 
8234e1e
f8903fb
 
 
 
 
 
 
 
 
 
 
db6e8b5
f8903fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3d92921
 
 
f8903fb
 
320a6a7
f8903fb
 
3d92921
f8903fb
 
 
 
 
 
 
 
 
 
c5232e0
f8903fb
 
 
 
 
 
 
 
db6e8b5
f8903fb
 
 
 
 
 
 
 
db6e8b5
f8903fb
 
 
 
 
3d92921
 
 
f8903fb
db6e8b5
 
 
 
 
 
 
 
 
 
f8903fb
db6e8b5
f8903fb
 
3d92921
f8903fb
3d92921
 
f8903fb
3d92921
 
 
f8903fb
 
 
db6e8b5
f8903fb
 
db6e8b5
f8903fb
 
 
db6e8b5
f8903fb
 
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
---
language: lg
datasets:
- common_voice
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Large Luganda by Lucio
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice lg
      type: common_voice
      args: lg
    metrics:
       - name: Test WER
         type: wer
         value: 22.82
---

# Wav2Vec2-Large-XLSR-53-lg

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Luganda using the [Common Voice](https://huggingface.co/datasets/common_voice) dataset, using train, validation and other (if the example had more upvotes than downvotes), and taking the test data for validation as well as test.
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", "lg", split="test[:2%]") 

processor = Wav2Vec2Processor.from_pretrained("lucio/wav2vec2-large-xlsr-luganda") 
model = Wav2Vec2ForCTC.from_pretrained("lucio/wav2vec2-large-xlsr-luganda")

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[:2]["speech"], 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 Luganda test data of Common Voice. (Available in Colab [here](https://colab.research.google.com/drive/1XxZ3mJOEXwIn-QH3C23jD_Qpom9aA1vH?usp=sharing).)


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

test_dataset = load_dataset("common_voice", "lg", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("lucio/wav2vec2-large-xlsr-luganda") 
model = Wav2Vec2ForCTC.from_pretrained("lucio/wav2vec2-large-xlsr-luganda")
model.to("cuda")

chars_to_ignore_regex = '[\[\],?.!;:%"“”(){}‟ˮʺ″«»/…‽�–]'
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

def remove_special_characters(batch):
    # word-internal apostrophes are marking contractions
    batch["norm_text"] = re.sub(r'[‘’´`]', r"'", batch["sentence"])
    # most other punctuation is ignored
    batch["norm_text"] = re.sub(chars_to_ignore_regex, "", batch["norm_text"]).lower().strip()
    batch["norm_text"] = re.sub(r"(-|' | '|  +)", " ", batch["norm_text"])
    # remove accents from a few characters (from loanwords, not tones)
    batch["norm_text"] = unidecode.unidecode(batch["norm_text"])
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
test_dataset = test_dataset.map(remove_special_characters)

def evaluate(batch):
    inputs = processor(batch["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)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["norm_text"])))
```

**Test Result**: 22.82 % 

## Training

The Common Voice `train`, `validation` and `other` datasets were used for training, augmented to twice the original size with added noise and manipulated pitch, phase and intensity.

The script used for training was just the `run_finetuning.py` script provided in OVHcloud's databuzzword/hf-wav2vec image.