Fine-tune Whisper-small for Korean Speech Recognition sample data (PoC)

Fine-tuning was performed using sample voices recorded from this csv data(https://github.com/hyeonsangjeon/job-transcribe/blob/main/meta_voice_data_3922.csv). We do not publish sample voices, so if you want to fine-tune yourself from scratch, please record separately or use a public dataset.

Fine tuning training based on the guide at https://huggingface.co/blog/fine-tune-whisper

[Note] In the voice recording data used for training, the speaker spoke clearly and slowly as if reading a textbook.

Training

Base model

OpenAI's whisper-small (https://huggingface.co/openai/whisper-small)

Parameters

We used heuristic parameters without separate hyperparameter tuning. The sampling rate is set to 16,000Hz.

learning_rate = 2e-5
epochs = 5
gradient_accumulation_steps = 4
per_device_train_batch_size = 4
fp16 = True
gradient_checkpointing = True
generation_max_length = 225

Usage

You need to install librosa package in order to convert wave to Mel Spectrogram. (pip install librosa)

inference.py

import librosa
import torch
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# prepare your sample data (.wav)
file = "nlp-voice-3922/data/0002d3428f0ddfa5a48eec5cc351daa8.wav"

# Convert to Mel Spectrogram
arr, sampling_rate = librosa.load(file, sr=16000)

# Load whisper model and processor
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("daekeun-ml/whisper-small-ko-finetuned-single-speaker-3922samples")

# Preprocessing
input_features = processor(arr, return_tensors="pt", sampling_rate=sampling_rate).input_features 

# Prediction
forced_decoder_ids = processor.get_decoder_prompt_ids(language="ko", task="transcribe")
predicted_ids = model.generate(input_features, forced_decoder_ids=forced_decoder_ids)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)

print(transcription)