Edit model card

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. αž¬αžŸαŸ’αžŸαžΈ αžͺαž‘αžΉαž€
    • 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
Downloads last month
148
Safetensors
Model size
316M params
Tensor type
F32
Β·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Spaces using vitouphy/wav2vec2-xls-r-300m-khmer 32

Evaluation results