w2v-bert-2.0-luganda-CV-train-validation-7.0

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the Luganda mozilla common voices 7.0 dataset. We use the train and validation set for training and the test set for evaluation. When using this dataset, make sure that the audio has a sampling rate of 16kHz.It achieves the following results on the test set:

• Loss: 0.2282
• Wer: 0.1933

Training and evaluation data

The model was trained on version 7 of the Luganda dataset of Mozilla common voices dataset. We used the train and validation set for training and the test dataset for validation. The training script was adapted from this transformers repo.

Training procedure

We trained the model on a 32 GB V100 GPU for 10 epochs using a learning rate of 5e-05. We used the AdamW optimizer.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.1+cu121
• Datasets 2.17.0
• Tokenizers 0.15.2

Usage

import torch
import torchaudio
from datasets import load_dataset
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor

test_dataset = load_dataset("common_voice", "lg", split="test[:10]")

model = AutoModelForCTC.from_pretrained("dmusingu/w2v-bert-2.0-luganda-CV-train-validation-7.0")
processor = Wav2Vec2BertProcessor.from_pretrained("dmusingu/w2v-bert-2.0-luganda-CV-train-validation-7.0")

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):
    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["sentence"][:2])

Evaluation

The model can be evaluated as follows on the Luganda test dataset.

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor
import re

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

model = AutoModelForCTC.from_pretrained("dmusingu/w2v-bert-2.0-luganda-CV-train-validation-7.0").to('cuda')
processor = Wav2Vec2BertProcessor.from_pretrained("dmusingu/w2v-bert-2.0-luganda-CV-train-validation-7.0")

chars_to_remove_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\»\«]'

test_dataset = test_dataset.cast_column("audio", Audio(sampling_rate=16_000))

def remove_special_characters(batch):
    # remove special characters
    batch["sentence"] = re.sub(chars_to_remove_regex, '', batch["sentence"]).lower()

    return batch

test_dataset = test_dataset.map(remove_special_characters)

def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
    batch["input_length"] = len(batch["input_features"])

    batch["labels"] = processor(text=batch["sentence"]).input_ids
    return batch

test_dataset = test_dataset.map(prepare_dataset, remove_columns=test_dataset.column_names)

# Evaluation is carried out with a batch size of 1
def map_to_result(batch):
  with torch.no_grad():
    input_values = torch.tensor(batch["input_features"], device="cuda").unsqueeze(0)
    logits = model(input_values).logits

  pred_ids = torch.argmax(logits, dim=-1)
  batch["pred_str"] = processor.batch_decode(pred_ids)[0]
  batch["text"] = processor.decode(batch["labels"], group_tokens=False)

  return batch

results = test_dataset.map(map_to_result)

print("Test WER: {:.3f}".format(wer_metric.compute(predictions=results["pred_str"], references=results["text"])))

Test Result: 19.33%