Vietnamese Speech Recognition using Wav2vec 2.0

Table of contents

1. Model Description
2. Implementation
3. Benchmark Result
4. Example Usage
5. Evaluation
6. Citation
7. Contact

Model Description

Fine-tuned the Wav2vec2-based model on about 160 hours of Vietnamese speech dataset from different resources, including VIOS, COMMON VOICE, FOSD and VLSP 100h. We have not yet incorporated the Language Model into our ASR system but still gained a promising result.

Implementation

We also provide code for Pre-training and Fine-tuning the Wav2vec2 model. If you wish to train on your dataset, check it out here:

• Pre-train code
• Fine-tune code

Benchmark WER Result

	VIVOS	COMMON VOICE 8.0
without LM	15.05	10.78
with LM	in progress	in progress

Example Usage

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import librosa
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

processor = Wav2Vec2Processor.from_pretrained("khanhld/wav2vec2-base-vietnamese-160h")
model = Wav2Vec2ForCTC.from_pretrained("khanhld/wav2vec2-base-vietnamese-160h")
model.to(device)

def transcribe(wav):
  input_values = processor(wav, sampling_rate=16000, return_tensors="pt").input_values
  logits = model(input_values.to(device)).logits
  pred_ids = torch.argmax(logits, dim=-1)
  pred_transcript = processor.batch_decode(pred_ids)[0]
  return pred_transcript
  
wav, _ = librosa.load('path/to/your/audio/file', sr = 16000)
print(f"transcript: {transcribe(wav)}")

Evaluation

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torch
import re
from datasets import load_dataset, load_metric, Audio

wer = load_metric("wer")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# load processor and model
processor = Wav2Vec2Processor.from_pretrained("khanhld/wav2vec2-base-vietnamese-160h")
model = Wav2Vec2ForCTC.from_pretrained("khanhld/wav2vec2-base-vietnamese-160h")
model.to(device)
model.eval()

# Load dataset
test_dataset = load_dataset("mozilla-foundation/common_voice_8_0", "vi", split="test", use_auth_token="your_huggingface_auth_token")
test_dataset = test_dataset.cast_column("audio", Audio(sampling_rate=16000))
chars_to_ignore = r'[,?.!\-;:"“%\'�]' # ignore special characters

# preprocess data
def preprocess(batch):
  audio = batch["audio"]
  batch["input_values"] = audio["array"]
  batch["transcript"] = re.sub(chars_to_ignore, '', batch["sentence"]).lower()
  return batch
  
# run inference
def inference(batch):
  input_values = processor(batch["input_values"], 
                            sampling_rate=16000, 
                            return_tensors="pt").input_values
  logits = model(input_values.to(device)).logits
  pred_ids = torch.argmax(logits, dim=-1)
  batch["pred_transcript"] = processor.batch_decode(pred_ids) 
  return batch
  
test_dataset = test_dataset.map(preprocess)
result = test_dataset.map(inference, batched=True, batch_size=1)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_transcript"], references=result["transcript"])))

Test Result: 10.78%

Citation

BibTeX

@mics{Duy_Khanh_Finetune_Wav2vec_2_0_2022,
  author = {Duy Khanh, Le},
  doi = {10.5281/zenodo.6542357},
  license = {CC-BY-NC-4.0},
  month = {5},
  title = {{Finetune Wav2vec 2.0 For Vietnamese Speech Recognition}},
  url = {https://github.com/khanld/ASR-Wa2vec-Finetune},
  year = {2022}
}

APA

Duy Khanh, L. (2022). Finetune Wav2vec 2.0 For Vietnamese Speech Recognition [Data set]. https://doi.org/10.5281/zenodo.6542357

Contact

• khanhld218@uef.edu.vn
• 
•