Update README.md
Browse files
README.md
CHANGED
@@ -62,17 +62,59 @@ It achieves the following results on the evaluation set:
|
|
62 |
- Since this dataset is small (4 hours of voice recording), we decided not to train that for too long to avoid overfitting and under-generalization.
|
63 |
- This model performs worse than its 300M-variant. Probably, we don't explore the hyper-parameter enough?
|
64 |
|
65 |
-
## Model description
|
66 |
|
67 |
-
|
|
|
|
|
|
|
|
|
|
|
68 |
|
69 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
|
73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
-
More information needed
|
76 |
|
77 |
## Training procedure
|
78 |
|
|
|
62 |
- Since this dataset is small (4 hours of voice recording), we decided not to train that for too long to avoid overfitting and under-generalization.
|
63 |
- This model performs worse than its 300M-variant. Probably, we don't explore the hyper-parameter enough?
|
64 |
|
|
|
65 |
|
66 |
+
## Installation
|
67 |
+
Install the following libraries on top of HuggingFace Transformers for the supports of language model.
|
68 |
+
```
|
69 |
+
pip install pyctcdecode
|
70 |
+
pip install https://github.com/kpu/kenlm/archive/master.zip
|
71 |
+
```
|
72 |
|
73 |
+
## Usage
|
74 |
+
|
75 |
+
**Approach 1:** Using HuggingFace's pipeline, this will cover everything end-to-end from raw audio input to text output.
|
76 |
+
```python
|
77 |
+
from transformers import pipeline
|
78 |
+
|
79 |
+
# Load the model
|
80 |
+
pipe = pipeline(model="vitouphy/wav2vec2-xls-r-300m-khmer")
|
81 |
+
|
82 |
+
# Process raw audio
|
83 |
+
output = pipe("sound_file.wav", chunk_length_s=10, stride_length_s=(4, 2))
|
84 |
+
```
|
85 |
+
|
86 |
+
**Approach 2:** More custom way to predict phonemes.
|
87 |
+
```python
|
88 |
+
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
|
89 |
+
import librosa
|
90 |
+
import torch
|
91 |
|
92 |
+
# load model and processor
|
93 |
+
processor = Wav2Vec2Processor.from_pretrained("vitouphy/wav2vec2-xls-r-300m-khmer")
|
94 |
+
model = Wav2Vec2ForCTC.from_pretrained("vitouphy/wav2vec2-xls-r-300m-khmer")
|
95 |
+
|
96 |
+
# Read and process the input
|
97 |
+
speech_array, sampling_rate = librosa.load("sound_file.wav", sr=16_000)
|
98 |
+
inputs = processor(speech_array, sampling_rate=16_000, return_tensors="pt", padding=True)
|
99 |
+
|
100 |
+
with torch.no_grad():
|
101 |
+
logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
|
102 |
+
|
103 |
+
predicted_ids = torch.argmax(logits, axis=-1)
|
104 |
+
predicted_sentences = processor.batch_decode(predicted_ids)
|
105 |
+
print(predicted_sentences)
|
106 |
+
```
|
107 |
+
|
108 |
+
## Intended uses & limitations
|
109 |
|
110 |
+
The data used for this model is only around 4 hours of recordings.
|
111 |
+
- We split into 80/10/10. Hence, the training hour is 3.2 hours, which is very very small.
|
112 |
+
- Yet, its performance is not too bad. Quite interesting for such small dataset, actually. You can try it out.
|
113 |
+
- Its limitation is:
|
114 |
+
- Rare characters, e.g. ឬស្សី ឪឡឹក
|
115 |
+
- Speech needs to be clear and articulate.
|
116 |
+
- More data to cover more vocabulary and character may help improve this system.
|
117 |
|
|
|
118 |
|
119 |
## Training procedure
|
120 |
|