Wav2vec2 Xlsr Korean Senior

🚀 wav2vec2-xlsr-korean-senior

This model is further fine-tuned from fleek/wav2vec-large-xlsr-korean using the AIhub Free Conversation Speech (Elderly Men and Women), aiming to provide more accurate automatic speech recognition for elderly Korean speakers.

🚀 Quick Start

When using this model, make sure that your speech input is sampled at 16kHz. The script used for training can be found here: https://github.com/hyyoka/wav2vec2-korean-senior

✨ Features

• Total train data size: 808,642
• Total valid data size: 159,970

💻 Usage Examples

Basic Usage

import torchaudio
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import re

def clean_up(transcription):
    hangul = re.compile('[^ ㄱ-ㅣ가-힣]+')
    result = hangul.sub('', transcription)
    return result

model_name = "hyyoka/wav2vec2-xlsr-korean-senior"
processor = Wav2Vec2Processor.from_pretrained(model_name)
model = Wav2Vec2ForCTC.from_pretrained(model_name)
speech_array, sampling_rate = torchaudio.load(wav_file)
feat = processor(speech_array[0], 
                            sampling_rate=16000, 
                            padding=True,
                            max_length=800000, 
                            truncation=True,
                            return_attention_mask=True,
                            return_tensors="pt",
                            pad_token_id=49
                            )
input = {'input_values': feat['input_values'],'attention_mask':feat['attention_mask']}

outputs = model(**input, output_attentions=True)
logits = outputs.logits
predicted_ids = logits.argmax(axis=-1)
transcription = processor.decode(predicted_ids[0])
stt_result = clean_up(transcription)

📄 License

This model is released under the Apache 2.0 license.