Create app.py

app.py

@@ -0,0 +1,91 @@
+import os, uuid, subprocess
+from flask import Flask, request, jsonify, render_template
+from faster_whisper import WhisperModel
+from transformers import pipeline, AutoTokenizer, AutoModel
+from deep_translator import GoogleTranslator
+import torch
+import torch.nn.functional as F
+
+app = Flask(__name__)
+
+# ---------------- LOAD MODELS ----------------
+print("Loading models...")
+
+whisper_model = WhisperModel("small", device="cpu", compute_type="int8")
+
+bart = pipeline("text2text-generation", model="facebook/bart-large")
+
+tokenizer = AutoTokenizer.from_pretrained("nlpie/clinical-roberta-base")
+roberta = AutoModel.from_pretrained("nlpie/clinical-roberta-base")
+
+# ---------------- EMBEDDING ----------------
+def get_embedding(text):
+    inputs = tokenizer(text, return_tensors="pt", truncation=True)
+    outputs = roberta(**inputs)
+    return outputs.last_hidden_state.mean(dim=1)
+
+# ---------------- REFERENCE ----------------
+ref_texts = {
+    "emergency": "severe chest pain unconscious breathing failure stroke",
+    "urgent": "fever infection moderate pain vomiting",
+    "non-urgent": "mild headache cold stable"
+}
+
+ref_embeddings = {k: get_embedding(v) for k, v in ref_texts.items()}
+
+# ---------------- ROUTES ----------------
+@app.route("/")
+def home():
+    return render_template("index.html")
+
+@app.route("/process", methods=["POST"])
+def process():
+
+    file = request.files["audio"]
+
+    uid = str(uuid.uuid4())
+    input_path = f"{uid}.webm"
+    wav_path = f"{uid}.wav"
+
+    file.save(input_path)
+
+    subprocess.run([
+        "ffmpeg", "-y", "-i", input_path,
+        "-ac", "1", "-ar", "16000", wav_path
+    ])
+
+    # Whisper
+    segments, info = whisper_model.transcribe(wav_path)
+    text = " ".join([seg.text for seg in segments])
+
+    # Translate
+    try:
+        text_en = GoogleTranslator(source="auto", target="en").translate(text)
+    except:
+        text_en = text
+
+    # BART clean
+    cleaned = bart(f"Extract symptoms: {text_en}")[0]["generated_text"]
+
+    # RoBERTa classify
+    emb = get_embedding(cleaned)
+
+    scores = {}
+    for k, ref in ref_embeddings.items():
+        scores[k] = F.cosine_similarity(emb, ref).item()
+
+    label = max(scores, key=scores.get)
+    confidence = scores[label]
+
+    score_map = {"emergency": 90, "urgent": 60, "non-urgent": 20}
+    score = int(score_map[label] * max(confidence, 0.5))
+
+    return jsonify({
+        "text": text,
+        "cleaned": cleaned,
+        "severity": label,
+        "score": score
+    })
+
+if __name__ == "__main__":
+    app.run(host="0.0.0.0", port=7860)