app.py

"""Minimum code to run OpenAI Whisper on Hugging Face (incl. ZeroGPU)"""
from functools import lru_cache
import whisper
import gradio as gr
import spaces

@lru_cache(maxsize=1)
def load_model():
    """Returns a newly loaded or cached whisper base model"""
    model = whisper.load_model("base")
    return model

@spaces.GPU(duration=120)
def transcribe(file_path: str):
    """Transcribes audio

    Keyword arguments:
    file_path -- path to audio file, passed from Gradio
    """
    model = load_model()
    result = model.transcribe(file_path)
    return result["text"]

interface = gr.Interface(
    fn = transcribe,
    inputs = gr.Audio(type="filepath", label="Upload audio"),
    outputs = gr.Textbox(label="Transcription Output"),
    title = "Minimum OpenAI Whisper",
    description = """
        <p>
            This repository is the most minimal example of a working OpenAI Whisper pipeline using ZeroGPU, and only ZeroGPU <sup><a href="#fn1" style="text-decoration: none;">[1]</a></sup>.
        </p>
        <p>
        Per the Unix philosophy: <q>do one thing and do it well</q>.
        </p>
        <p id="fn1">
            <small>[1] Because the build is broken on paid hardware, something about a Gradio version not existing, PRs welcome</small>
        </p>
    """
)

interface.queue().launch()