app.py

import os
from flask import Flask, request, jsonify, render_template
from transformers import pipeline
from flask_cors import CORS
from pydub import AudioSegment
from io import BytesIO
import Levenshtein

# Set the FFmpeg paths explicitly
AudioSegment.converter = "/usr/bin/ffmpeg"
AudioSegment.ffprobe = "/usr/bin/ffprobe"

# Set Hugging Face cache directory to avoid permission issues
os.environ['HF_HOME'] = '/tmp/.cache'

app = Flask(__name__)
CORS(app)

# Use Hugging Face ASR pipeline for automatic speech recognition
asr_pipeline = pipeline("automatic-speech-recognition", model="jonatasgrosman/wav2vec2-large-xlsr-53-arabic")


def convert_to_wav(audio_bytes):
    """Convert audio bytes to wav format using pydub"""
    try:
        audio = AudioSegment.from_file(BytesIO(audio_bytes))  # Auto-detect format
        wav_io = BytesIO()
        audio.export(wav_io, format="wav")
        wav_io.seek(0)
        return wav_io
    except Exception as e:
        print(f"Error converting audio: {e}")
        return None


def transcribe_audio(audio_bytes):
    """Transcribes the audio using the Hugging Face ASR pipeline."""
    wav_io = convert_to_wav(audio_bytes)
    if wav_io is None:
        raise Exception("Could not convert audio to WAV format")

    # Read the audio file into bytes for the ASR pipeline
    wav_io.seek(0)
    transcription = asr_pipeline(wav_io)["text"]
    return transcription.strip()


def levenshtein_similarity(transcription1, transcription2):
    distance = Levenshtein.distance(transcription1, transcription2)
    max_len = max(len(transcription1), len(transcription2))
    return 1 - distance / max_len


@app.route('/')
def index():
    return render_template('index.html')


@app.route('/transcribe', methods=['POST'])
def transcribe():
    original_audio = request.files['original_audio']
    user_audio = request.files['user_audio']

    original_audio_bytes = original_audio.read()
    user_audio_bytes = user_audio.read()

    try:
        transcription_original = transcribe_audio(original_audio_bytes)
        transcription_user = transcribe_audio(user_audio_bytes)
    except Exception as e:
        return jsonify({"error": str(e)}), 500

    similarity_score = levenshtein_similarity(transcription_original, transcription_user)

    return jsonify({
        "transcription_original": transcription_original,
        "transcription_user": transcription_user,
        "similarity_score": similarity_score
    })


if __name__ == '__main__':
    app.run(debug=False, port=7860, host='0.0.0.0')