app.py

import os
os.system("pip install git+https://github.com/sanchit-gandhi/whisper-jax.git")

# import whisper
from flask import Flask, jsonify, request
import requests
import time
# from transformers import pipeline
from whisper_jax import FlaxWhisperPipline
import jax.numpy as jnp



# model = whisper.load_model("small")
# pipe = pipeline(
#     "automatic-speech-recognition",
#     model="openai/whisper-small.en",
#     chunk_length_s=15,
#     device=model.device,
# )
pipe = FlaxWhisperPipline("openai/whisper-small", dtype=jnp.bfloat16, batch_size=16)

app = Flask(__name__)
app.config['TIMEOUT'] = 60 * 10 # 10 mins

@app.route("/")
def indexApi():
    return jsonify({"output": "okay"})

@app.route("/run", methods=['POST'])
def runApi():
    start_time = time.time()

    audio_url = request.form.get("audio_url")

    response = requests.get(audio_url)

    if response.status_code == requests.codes.ok:
        with open("audio.mp3", "wb") as f:
            f.write(response.content)
      
    else:
        return jsonify({
            "result": "Unable to save file, status code:  {response.status_code}" ,
        }), 400

    audio = "audio.mp3"

    # audioOri = whisper.load_audio(audio)
    # audio = whisper.pad_or_trim(audioOri)
    
    # mel = whisper.log_mel_spectrogram(audio).to(model.device)
    
    # _, probs = model.detect_language(mel)
    
    # options = whisper.DecodingOptions(fp16 = False)
    # result = whisper.decode(model, mel, options)

    # test 2
    # ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
    # sample = ds[0]["audio"]
    prediction = pipe(audio,  task="transcribe")["text"]
  

    end_time = time.time()
    total_time = end_time - start_time

    return jsonify({
        "audio_url": audio_url,
        "result": prediction,
        "exec_time_sec": total_time
    })

if __name__ == "__main__":
    app.run(host="0.0.0.0", port=7860)
        
# def inference(audio):
#     audio = whisper.load_audio(audio)
#     audio = whisper.pad_or_trim(audio)
    
#     mel = whisper.log_mel_spectrogram(audio).to(model.device)
    
#     _, probs = model.detect_language(mel)
    
#     options = whisper.DecodingOptions(fp16 = False)
#     result = whisper.decode(model, mel, options)
    
#     # print(result.text)
#     return result.text, gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)