app.py

import os
import time
import warnings
from pathlib import Path

import gradio as gr
import librosa
import spaces
import torch
from transformers import pipeline, WhisperConfig

warnings.filterwarnings("ignore")

is_hf = os.getenv("SYSTEM") == "spaces"
# reference from litagin / galgame-whisper-wip

generate_kwargs = {
    "max_new_tokens": 256,
}

pipe = pipeline(
    "automatic-speech-recognition",
    model="AkitoP/whisper-large-v3-japense-phone_accent",
    device="cuda" if torch.cuda.is_available() else "cpu",
)


@spaces.GPU
def transcribe(audio: str) -> str:
    result = pipe(audio, generate_kwargs=generate_kwargs)["text"]
    print(result)
    return result


initial_md = """
# Whisper Large V3 Japanese Phone Accent

A Whisper model fine-tuned to transcribe Japanese speech into Katakana with pitch accent annotations. Built on whisper-large-v3-turbo, it uses a subset (1/20) of the Galgame-Speech dataset and the jsut-5000 dataset.
"""

with gr.Blocks() as app:
    gr.Markdown(initial_md)
    audio = gr.Audio(type="filepath")
    transcribe_btn = gr.Button("Transcribe")
    output = gr.Textbox(label="Result")
    transcribe_btn.click(fn=transcribe,inputs=[audio], outputs=[output])
    
app.launch(inbrowser=True)