This project involves fine-tuning the Whisper-V3-Turbo model to improve its performance for Automatic Speech Recognition (ASR) in the Vietnamese language. The model was trained for 240 hours using a single Nvidia A6000 GPU.

Data Sources

The training data comes from various Vietnamese speech corpora. Below is a list of datasets used for training:

capleaf/viVoice
NhutP/VSV-1100
doof-ferb/fpt_fosd
doof-ferb/infore1_25hours
google/fleurs (vi_vn)
doof-ferb/LSVSC
quocanh34/viet_vlsp
linhtran92/viet_youtube_asr_corpus_v2
doof-ferb/infore2_audiobooks
linhtran92/viet_bud500

Model

The model used in this project is the Whisper-V3-Turbo. Whisper is a multilingual ASR model trained on a large and diverse dataset. The version used here has been fine-tuned specifically for the Vietnamese language.

Training Configuration

GPU Used: Nvidia A6000
Training Time: 240 hours
Wandb report

Usage

To use the fine-tuned model, follow the steps below:

python

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "suzii/vi-whisper-large-v3-turbo-v1"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)
result = pipe("your-audio.mp3", return_timestamps=True)

Acknowledgements

This project would not be possible without the following datasets: