Hugging Face's logo

nguyenvulebinh
/
wav2vec2-large-vi

Transformers
PyTorch
Vietnamese
wav2vec2
pretraining
speech
Inference Endpoints
Model card Files Community
1
wav2vec2-large-vi / README.md
nguyenvulebinh's picture
nguyenvulebinh
Update README.md
3a205ff
preview code
|
raw
history blame
3.54 kB
metadata
language: vi
datasets:
  - youtube-vi-13k-hours
tags:
  - speech
license: cc-by-nc-4.0

Vietnamese Self-Supervised Learning Wav2Vec2 model

Model

We use wav2vec2 architecture for doing Self-Supervised learning

Data

Our self-supervised model is pre-trained on a massive audio set of 13k hours of Vietnamese youtube audio, which includes:

  • Clean audio
  • Noise audio
  • Conversation
  • Multi-gender and dialects

Download

We have already upload our pre-trained model to the Huggingface. The base model trained 35 epochs and the large model trained 20 epochs in about 30 days using TPU V3-8.

Usage 

from transformers import Wav2Vec2ForPreTraining, Wav2Vec2Processor

model_name = 'nguyenvulebinh/wav2vec2-base-vi'
# model_name = 'nguyenvulebinh/wav2vec2-large-vi'

model = Wav2Vec2ForPreTraining.from_pretrained(model_name)
processor = Wav2Vec2Processor.from_pretrained(model_name)

Since our model has the same architecture as the English wav2vec2 version, you can use this notebook for more information on how to fine-tune the model.

Finetuned version

VLSP 2020 ASR dataset

Benchmark WER result on VLSP T1 testset:

base model large model
without LM 8.66 6.90
with 5-grams LM 6.53 5.32

Usage

#pytorch
#!pip install transformers==4.20.0
#!pip install https://github.com/kpu/kenlm/archive/master.zip
#!pip install pyctcdecode==0.4.0
from transformers.file_utils import cached_path, hf_bucket_url
from importlib.machinery import SourceFileLoader
from transformers import Wav2Vec2ProcessorWithLM
from IPython.lib.display import Audio
import torchaudio
import torch

# Load model & processor
model_name = "nguyenvulebinh/wav2vec2-base-vi-vlsp2020"
# model_name = "nguyenvulebinh/wav2vec2-large-vi-vlsp2020"
model = SourceFileLoader("model", cached_path(hf_bucket_url(model_name,filename="model_handling.py"))).load_module().Wav2Vec2ForCTC.from_pretrained(model_name)
processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_name)

# Load an example audio (16k)
audio, sample_rate = torchaudio.load(cached_path(hf_bucket_url(model_name, filename="t2_0000006682.wav")))
input_data = processor.feature_extractor(audio[0], sampling_rate=16000, return_tensors='pt')

# Infer
output = model(**input_data)

# Output transcript without LM
print(processor.tokenizer.decode(output.logits.argmax(dim=-1)[0].detach().cpu().numpy()))

# Output transcript with LM
print(processor.decode(output.logits.cpu().detach().numpy()[0], beam_width=100).text)

Acknowledgment

  • We would like to thank the Google TPU Research Cloud (TRC) program and Soonson Kwon (Google ML Ecosystem programs Lead) for their support.
  • Special thanks to my colleagues at VietAI and VAIS for their advice.

Contact

nguyenvulebinh@gmail.com / binh@vietai.org

Follow