vitouphy/wav2vec2-xls-r-1b-khmer

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the openslr dataset. It achieves the following results on the evaluation set:

• Loss: 0.4239
• Wer: 0.4221

Evaluation results on OpenSLR "test" (self-split 10%) (Running ./eval.py):

• WER: 0.4490281634272114
• CER: 0.12198285179047481

Evaluation results on OpenSLR "test" with LM ngram (self-split 10%) (Running ./eval.py):

• WER: 0.32130107100357
• CER: 0.09345053678218891

Note

• 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.
• This model performs worse than its 300M-variant. Probably, we don't explore the hyper-parameter enough?

Installation

Install the following libraries on top of HuggingFace Transformers for the supports of language model.

pip install pyctcdecode
pip install https://github.com/kpu/kenlm/archive/master.zip

Usage

Approach 1: Using HuggingFace's pipeline, this will cover everything end-to-end from raw audio input to text output.

from transformers import pipeline

# Load the model
pipe = pipeline(model="vitouphy/wav2vec2-xls-r-300m-khmer")

# Process raw audio
output = pipe("sound_file.wav", chunk_length_s=10, stride_length_s=(4, 2))

Approach 2: More custom way to predict phonemes.

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC 
import librosa
import torch

# load model and processor
processor = Wav2Vec2Processor.from_pretrained("vitouphy/wav2vec2-xls-r-300m-khmer")
model = Wav2Vec2ForCTC.from_pretrained("vitouphy/wav2vec2-xls-r-300m-khmer")

# Read and process the input
speech_array, sampling_rate = librosa.load("sound_file.wav", sr=16_000)
inputs = processor(speech_array, 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, axis=-1)      
predicted_sentences = processor.batch_decode(predicted_ids)
print(predicted_sentences)

Intended uses & limitations

The data used for this model is only around 4 hours of recordings.

• We split into 80/10/10. Hence, the training hour is 3.2 hours, which is very very small.
• Yet, its performance is not too bad. Quite interesting for such small dataset, actually. You can try it out.
• Its limitation is:
  • Rare characters, e.g. ឬស្សី ឪឡឹក
  • Speech needs to be clear and articulate.
• More data to cover more vocabulary and character may help improve this system.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 75
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.5671	5.47	400	12.0218	1.0
3.5159	10.95	800	10.6337	1.0
2.4543	16.43	1200	1.8256	0.9839
1.9437	21.91	1600	1.1237	0.9173
1.696	27.39	2000	0.8246	0.7700
1.5342	32.87	2400	0.6433	0.6594
1.4509	38.35	2800	0.5500	0.5787
1.3478	43.83	3200	0.5070	0.4907
1.3096	49.31	3600	0.4692	0.4726
1.2532	54.79	4000	0.4448	0.4479
1.2291	60.27	4400	0.4374	0.4366
1.196	65.75	4800	0.4314	0.4310
1.1862	71.23	5200	0.4239	0.4221

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0