vitouphy/wav2vec2-xls-r-300m-khmer

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the openslr dataset. It achieves the following results on the evaluation set:

• Loss: 0.3281
• Wer: 0.3462

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

• WER: 0.3216977389924633
• CER: 0.08653361193169537

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

• WER: 0.257040856802856
• CER: 0.07025001801282513

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: 5e-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: 1000
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.0795	5.47	400	4.4121	1.0
3.5658	10.95	800	3.5203	1.0
3.3689	16.43	1200	2.8984	0.9996
2.01	21.91	1600	1.0041	0.7288
1.6783	27.39	2000	0.6941	0.5989
1.527	32.87	2400	0.5599	0.5282
1.4278	38.35	2800	0.4827	0.4806
1.3458	43.83	3200	0.4429	0.4532
1.2893	49.31	3600	0.4156	0.4330
1.2441	54.79	4000	0.4020	0.4040
1.188	60.27	4400	0.3777	0.3866
1.1628	65.75	4800	0.3607	0.3858
1.1324	71.23	5200	0.3534	0.3604
1.0969	76.71	5600	0.3428	0.3624
1.0897	82.19	6000	0.3387	0.3567
1.0625	87.66	6400	0.3339	0.3499
1.0601	93.15	6800	0.3288	0.3446
1.0474	98.62	7200	0.3281	0.3462

Framework versions

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