Edit model card

Vietnamese end-to-end speech recognition using wav2vec 2.0



Facebook's Wav2Vec2

Model description

Our models are pre-trained on 13k hours of Vietnamese youtube audio (un-label data) and fine-tuned on 250 hours labeled of VLSP ASR dataset on 16kHz sampled speech audio.

We use wav2vec2 architecture for the pre-trained model. Follow wav2vec2 paper:

For the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler.

For fine-tuning phase, wav2vec2 is fine-tuned using Connectionist Temporal Classification (CTC), which is an algorithm that is used to train neural networks for sequence-to-sequence problems and mainly in Automatic Speech Recognition and handwriting recognition.

Model #params Pre-training data Fine-tune data
base 95M 13k hours 250 hours

In a formal ASR system, two components are required: acoustic model and language model. Here ctc-wav2vec fine-tuned model works as an acoustic model. For the language model, we provide a 4-grams model trained on 2GB of spoken text.

Detail of training and fine-tuning process, the audience can follow fairseq github and huggingface blog.

Benchmark WER result:

without LM 10.77 18.34 13.33 51.45
with 4-grams LM 6.15 11.52 9.11 40.81

Example usage

When using the model make sure that your speech input is sampled at 16Khz. Audio length should be shorter than 10s. Following the Colab link below to use a combination of CTC-wav2vec and 4-grams LM.

Open In Colab

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import soundfile as sf
import torch

# load model and tokenizer
processor = Wav2Vec2Processor.from_pretrained("nguyenvulebinh/wav2vec2-base-vietnamese-250h")
model = Wav2Vec2ForCTC.from_pretrained("nguyenvulebinh/wav2vec2-base-vietnamese-250h")

# define function to read in sound file
def map_to_array(batch):
    speech, _ = sf.read(batch["file"])
    batch["speech"] = speech
    return batch

# load dummy dataset and read soundfiles
ds = map_to_array({
    "file": 'audio-test/t1_0001-00010.wav'

# tokenize
input_values = processor(ds["speech"], return_tensors="pt", padding="longest").input_values  # Batch size 1

# retrieve logits
logits = model(input_values).logits

# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)

Model Parameters License

The ASR model parameters are made available for non-commercial use only, under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. You can find details at: https://creativecommons.org/licenses/by-nc/4.0/legalcode



  author = {Thai Binh Nguyen},
  doi = {10.5281/zenodo.5356039},
  month = {09},
  title = {{Vietnamese end-to-end speech recognition using wav2vec 2.0}},
  url = {https://github.com/vietai/ASR},
  year = {2021}

Please CITE our repo when it is used to help produce published results or is incorporated into other software.


nguyenvulebinh@gmail.com / binh@vietai.org


Downloads last month
Hosted inference API
or or
This model can be loaded on the Inference API on-demand.

Dataset used to train nguyenvulebinh/wav2vec2-base-vietnamese-250h

Spaces using nguyenvulebinh/wav2vec2-base-vietnamese-250h

Evaluation results