NhutP/ViWhisper-medium

Introduction

• We release a new model for Vietnamese speech regconition task.
• We fine-tuned openai/whisper-medium on our new dataset VSV-1100.

Training data

VSV-1100	T2S*	CMV14-vi	VIVOS	VLSP2021	Total
1100 hours	11 hours	3.04 hours	13.94 hours	180 hours	1308 hours

* We use a text-to-speech model to generate sentences containing words that do not appear in our dataset.

WER result

CMV14-vi	VIVOS	VLSP2020-T1	VLSP2020-T2	VLSP2021-T1	VLSP2021-T2	Bud500
8.1	4.69	13.22	28.76	11.78	8.28	5.38

Usage

Inference

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import librosa
# load model and processor
processor = WhisperProcessor.from_pretrained("NhutP/ViWhisper-medium")
model = WhisperForConditionalGeneration.from_pretrained("NhutP/ViWhisper-medium")
model.config.forced_decoder_ids = None

# load a sample
array, sampling_rate = librosa.load('path_to_audio', sr = 16000) # Load some audio sample
input_features = processor(array, sampling_rate=sampling_rate, return_tensors="pt").input_features 
# generate token ids
predicted_ids = model.generate(input_features)
# decode token ids to text
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)

Use with pipeline

from transformers import pipeline
pipe = pipeline(
    "automatic-speech-recognition",
    model="NhutP/ViWhisper-medium",
    max_new_tokens=128,
    chunk_length_s=30,
    return_timestamps=False,
    device= '...' # 'cpu' or 'cuda'
) 
output = pipe(path_to_audio_samplingrate_16000)['text']

Citation

@misc{VSV-1100,
    author = {Pham Quang Nhut and Duong Pham Hoang Anh and Nguyen Vinh Tiep},
    title = {VSV-1100: Vietnamese social voice dataset},
    url = {https://github.com/NhutP/VSV-1100},
    year = {2024}
}

Also, please give us a star on github: https://github.com/NhutP/ViWhisper if you find our project useful

Contact me at: [email protected] (Pham Quang Nhut)