from fastapi import FastAPI, UploadFile, File
from transformers import pipeline, WhisperForConditionalGeneration, WhisperProcessor
import torch
import tempfile
import os
import time
from fastapi.responses import HTMLResponse
from fastapi.staticfiles import StaticFiles
# Define FastAPI app
app = FastAPI()
# Load the Whisper model once during startup
device = 0 if torch.cuda.is_available() else -1 # Use GPU if available, otherwise CPU
asr_pipeline = pipeline(model="openai/whisper-small", device=device) # Initialize Whisper model
# asr_pipeline = pipeline( model="openai/whisper-small", device=device, language="pt")
# Basic GET endpoint
@app.get("/")
def read_root():
return {"message": "Welcome to the FastAPI app on Hugging Face Spaces!"}
# POST endpoint to transcribe audio
@app.post("/transcribe/")
async def transcribe_audio(file: UploadFile = File(...)):
start_time = time.time()
# Save the uploaded file using a temporary file manager
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_audio_file:
temp_audio_file.write(await file.read())
temp_file_path = temp_audio_file.name
# Transcribe the audio with long-form generation enabled
transcription_start = time.time()
transcription = asr_pipeline(temp_file_path, return_timestamps=True) # Enable timestamp return for long audio files
transcription_end = time.time()
# Clean up temporary file after use
os.remove(temp_file_path)
# Log time durations
end_time = time.time()
print(f"Time to transcribe audio: {transcription_end - transcription_start:.4f} seconds")
print(f"Total execution time: {end_time - start_time:.4f} seconds")
return {"transcription": transcription['text']}
@app.get("/playground/", response_class=HTMLResponse)
def playground():
html_content = """
Voice Recorder
Record your voice
Press start to record your voice...
"""
return HTMLResponse(content=html_content)
# If running as the main module, start Uvicorn
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=7860)