File size: 4,518 Bytes
d47bd37
 
 
 
 
 
 
 
 
 
 
 
67d044b
d47bd37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67d044b
 
 
 
d47bd37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67d044b
 
 
 
d47bd37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a032e29
e69a778
a032e29
e69a778
d47bd37
 
 
 
 
 
 
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
138
139
140
141
---
language: lg
datasets:
- common_voice
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
license: apache-2.0
model-index:
- name: Wav2Vec2 Luganda by Indonesian-NLP
  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: 7.53
---

# Automatic Speech Recognition for Luganda

This is the model built for the 
[Mozilla Luganda Automatic Speech Recognition competition](https://zindi.africa/competitions/mozilla-luganda-automatic-speech-recognition).
It is a fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53)
model on the [Luganda Common Voice dataset](https://huggingface.co/datasets/common_voice) version 7.0.

We also provide a [live demo](https://huggingface.co/spaces/indonesian-nlp/luganda-asr) to test the model.

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("indonesian-nlp/wav2vec2-luganda")
model = Wav2Vec2ForCTC.from_pretrained("indonesian-nlp/wav2vec2-luganda")

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

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    if "audio" in batch:
        speech_array = torch.tensor(batch["audio"]["array"])
    else:
        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[:2]["sentence"])
```


## Evaluation

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

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

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

processor = Wav2Vec2Processor.from_pretrained("indonesian-nlp/wav2vec2-luganda")
model = Wav2Vec2ForCTC.from_pretrained("indonesian-nlp/wav2vec2-luganda") 
model.to("cuda")

chars_to_ignore = [",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�", "‘", "’", "’"]
chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'

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):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
    if "audio" in batch:
        speech_array = torch.tensor(batch["audio"]["array"])
    else:
        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)

# Preprocessing the datasets.
# We need to read the audio files as arrays
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["sentence"])))
```

WER without KenLM: 15.38 %

WER With KenLM:

**Test Result**: 7.53 %

## Training

The Common Voice `train`, `validation`, and ... datasets were used for training as well as ... and ...  # TODO

The script used for training can be found [here](https://github.com/indonesian-nlp/luganda-asr)