app.py

import gradio as gr
import numpy as np

from librosa import resample
from transformers import pipeline

pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base.en", chunk_length_s=30)

def transcribe(audio_in):
  orig_sr, samples = audio_in
  min_s, max_s = min(samples), max(samples)
  range_in = (max_s - min_s)
  samples_scl = np.array(samples) / range_in
  min_scl = min_s / range_in
  samples_f = 2.0 * (samples_scl - min_scl) - 1.0
  resamples = resample(samples_f, orig_sr=orig_sr, target_sr=16000)
  prediction = pipe(resamples.copy(), batch_size=8)
  return prediction["text"].strip().lower()


with gr.Blocks() as demo:
  gr.Markdown("""
              # 5020 Audio Transcription.
              ## API for [whisper-base.en](https://huggingface.co/openai/whisper-base.en) english model\
              to help with Audio Analysis exercises.
              """)

  gr.Interface(
    transcribe,
    inputs=gr.Audio(type="numpy"),
    outputs="text",
    flagging_mode="never",
    cache_examples=True,
    examples=[["./audio/plain_01.wav"]]
  )

if __name__ == "__main__":
   demo.launch(ssr_mode=False)