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.
VIVOS | COMMON VOICE VI | VLSP-T1 | VLSP-T2 | |
---|---|---|---|---|
without LM | 10.77 | 18.34 | 13.33 | 51.45 |
with 4-grams LM | 6.15 | 11.52 | 9.11 | 40.81 |
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.
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)
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
@misc{Thai_Binh_Nguyen_wav2vec2_vi_2021,
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.