Wav2Vec2-Base-Vietnamese-270h

Fine-tuned Wav2Vec2 model on Vietnamese Speech Recognition task using about 270h labelled data combined from multiple datasets including Common Voice , VIVOS , VLSP2020 . The model was fine-tuned using SpeechBrain toolkit with a custom tokenizer. For a better experience, we encourage you to learn more about SpeechBrain . When using this model, make sure that your speech input is sampled at 16kHz. Please refer to huggingface blog or speechbrain on how to fine-tune Wav2Vec2 model on a specific language.

Benchmark WER result:

The language model was trained using OSCAR dataset on about 32GB of crawled text.

Install SpeechBrain

To use this model, you should install speechbrain > 0.5.10

Usage

The model can be used directly (without a language model) as follows:

from speechbrain.pretrained import EncoderASR

model = EncoderASR.from_hparams(source="dragonSwing/wav2vec2-base-vn-270h", savedir="pretrained_models/asr-wav2vec2-vi")
model.transcribe_file('dragonSwing/wav2vec2-base-vn-270h/example.mp3')
# Output: được hồ chí minh coi là một động lực lớn của sự phát triển đất nước

Inference on GPU

To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.

Evaluation

The model can be evaluated as follows on the Vietnamese test data of Common Voice 8.0.

import torch
import torchaudio
from datasets import load_dataset, load_metric, Audio
from transformers import Wav2Vec2FeatureExtractor
from speechbrain.pretrained import EncoderASR
import re
test_dataset = load_dataset("mozilla-foundation/common_voice_8_0", "vi", split="test", use_auth_token=True)
test_dataset = test_dataset.cast_column("audio", Audio(sampling_rate=16_000))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
wer = load_metric("wer")
extractor = Wav2Vec2FeatureExtractor.from_pretrained("dragonSwing/wav2vec2-base-vn-270h")
model = EncoderASR.from_hparams(source="dragonSwing/wav2vec2-base-vn-270h", savedir="pretrained_models/asr-wav2vec2-vi", run_opts={'device': device})
chars_to_ignore_regex = r'[,?.!\-;:"“%\'�]'
# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
  audio = batch["audio"]
  batch["target_text"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
  batch['speech'] = audio['array']
  return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)

def evaluate(batch):
  # For padding inputs only
  inputs = extractor(
    batch['speech'], 
    sampling_rate=16000, 
    return_tensors="pt", 
    padding=True, 
    do_normalize=False
  ).input_values
  input_lens = torch.ones(inputs.shape[0])
  pred_str, pred_tokens = model.transcribe_batch(inputs, input_lens)
  batch["pred_strings"] = pred_str
  
  return batch
result = test_dataset.map(evaluate, batched=True, batch_size=1)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["target_text"])))

Test Result : 12.155553%

@misc{SB2021,
    author = {Ravanelli, Mirco and Parcollet, Titouan and Rouhe, Aku and Plantinga, Peter and Rastorgueva, Elena and Lugosch, Loren and Dawalatabad, Nauman and Ju-Chieh, Chou and Heba, Abdel and Grondin, Francois and Aris, William and Liao, Chien-Feng and Cornell, Samuele and Yeh, Sung-Lin and Na, Hwidong and Gao, Yan and Fu, Szu-Wei and Subakan, Cem and De Mori, Renato and Bengio, Yoshua },
    title = {SpeechBrain},
    year = {2021},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\\\\url{https://github.com/speechbrain/speechbrain}},
  }

SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to be simple, extremely flexible, and user-friendly. Competitive or state-of-the-art performance is obtained in various domains. Website: https://speechbrain.github.io GitHub: https://github.com/speechbrain/speechbrain