app.py

ABOUT = """
# Fast Whisper Turbo ⚡

Ultra-fast Whisper V3 Turbo inference, with enhancements sourced from [insanely-fast-whisper](https://github.com/Vaibhavs10/insanely-fast-whisper).
"""

CREDITS = """
## Credits

This project was made possible through the work of several other projects:

- [insanely-fast-whisper](https://github.com/Vaibhavs10/insanely-fast-whisper)
"""

import spaces
import subprocess
subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
) # https://huggingface.co/spaces/zero-gpu-explorers/README/discussions/75#666e4681303f0a5d67175a90
import gradio as gr
from transformers import pipeline
import torch

pipe = pipeline(
    "automatic-speech-recognition",
    model="openai/whisper-large-v3-turbo",
    torch_dtype=torch.float16,
    device="cuda:0",
    model_kwargs={"attn_implementation": "flash_attention_2"},
)

@spaces.GPU
def transcribe(audio, task):
    gr.Info("Starting transcription task")
    outputs = pipe(
        audio,
        chunk_length_s=30,
        batch_size=128,
        generate_kwargs={"task": task},
        return_timestamps=False,
    )
    gr.Info("Finished transcription task")
    return outputs['text'].strip()

with gr.Blocks() as demo:
    gr.Markdown(ABOUT)
    audio = gr.Audio(label="Audio", type="filepath", interactive=True)
    task = gr.Radio(["transcribe", "translate"], label="Task", value="transcribe", interactive=True)
    btn = gr.Button("Transcribe", variant="primary")
    output = gr.Textbox(label="Transcription", interactive=False)
    btn.click(transcribe, inputs=[audio, task], outputs=output)
    gr.Markdown(CREDITS)

demo.queue().launch()