SONICS detect API (FastAPI wrapper, SpecTTTra-alpha-120s)

Dockerfile

@@ -0,0 +1,24 @@
+FROM python:3.11-slim
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends ffmpeg libsndfile1 git && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN useradd -m -u 1000 user
+USER user
+ENV HOME=/home/user \
+    PATH=/home/user/.local/bin:$PATH \
+    HF_HOME=/home/user/.cache/huggingface
+WORKDIR /home/user/app
+
+RUN pip install --no-cache-dir --user torch --index-url https://download.pytorch.org/whl/cpu && \
+    pip install --no-cache-dir --user librosa soundfile numpy fastapi "uvicorn[standard]" python-multipart huggingface_hub timm pandas && \
+    pip install --no-cache-dir --user git+https://github.com/awsaf49/sonics.git
+
+# Bake model weights into the image so cold starts skip the download.
+RUN python -c "from sonics import HFAudioClassifier; HFAudioClassifier.from_pretrained('awsaf49/sonics-spectttra-alpha-120s')"
+
+COPY --chown=user app.py .
+
+EXPOSE 7860
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -1,10 +1,20 @@
 ---
 title: Sonics Detect Api
-emoji: 🏢
-colorFrom: green
-colorTo: blue
+emoji: 🎵
+colorFrom: indigo
+colorTo: purple
 sdk: docker
+app_port: 7860
 pinned: false
+license: mit
+short_description: AI-generated song detection API (SONICS SpecTTTra)
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# SONICS Detect API
+
+Minimal REST wrapper around [SONICS SpecTTTra](https://huggingface.co/awsaf49/sonics-spectttra-alpha-120s)
+(ICLR 2025, MIT) for end-to-end AI-generated song detection.
+
+- `GET /` — health check
+- `POST /detect` — multipart form with an `audio` file, returns `{"ai_prob": 0..1}`.
+  Requires `X-Detect-Key` header when the `DETECT_API_KEY` secret is set.

app.py

@@ -0,0 +1,80 @@
+import os
+import tempfile
+
+import librosa
+import numpy as np
+import torch
+from fastapi import FastAPI, File, Header, HTTPException, UploadFile
+
+MODEL_ID = os.environ.get("MODEL_ID", "awsaf49/sonics-spectttra-alpha-120s")
+API_KEY = os.environ.get("DETECT_API_KEY", "")
+MAX_BYTES = 25 * 1024 * 1024
+
+torch.set_num_threads(2)
+
+app = FastAPI(title="SONICS Detect API")
+model = None
+
+
+@app.on_event("startup")
+def load_model():
+    global model
+    from sonics import HFAudioClassifier
+
+    m = HFAudioClassifier.from_pretrained(MODEL_ID)
+    m.eval()
+    model = m
+
+
+@app.get("/")
+def health():
+    return {"ok": True, "model": MODEL_ID, "loaded": model is not None}
+
+
+@app.post("/detect")
+def detect(audio: UploadFile = File(...), x_detect_key: str = Header(default="")):
+    if API_KEY and x_detect_key != API_KEY:
+        raise HTTPException(status_code=401, detail="Invalid detect key")
+    if model is None:
+        raise HTTPException(status_code=503, detail="Model still loading, retry shortly")
+
+    data = audio.file.read()
+    if not data:
+        raise HTTPException(status_code=400, detail="Empty file")
+    if len(data) > MAX_BYTES:
+        raise HTTPException(status_code=413, detail="File too large (max 25MB)")
+
+    suffix = os.path.splitext(audio.filename or "")[1] or ".mp3"
+    try:
+        with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+            tmp.write(data)
+            tmp_path = tmp.name
+        try:
+            y, sr = librosa.load(tmp_path, sr=16000, mono=True)
+        finally:
+            os.unlink(tmp_path)
+    except Exception:
+        raise HTTPException(status_code=400, detail="Could not decode audio")
+
+    if y.size < sr * 3:
+        raise HTTPException(status_code=400, detail="Audio too short (min 3s)")
+
+    # Same chunking as the official SONICS demo: score the middle max_time window.
+    max_time = model.config.audio.max_time
+    chunk_samples = int(max_time * sr)
+    total_chunks = len(y) // chunk_samples
+    middle_idx = total_chunks // 2
+    start = middle_idx * chunk_samples
+    chunk = y[start : start + chunk_samples]
+    if len(chunk) < chunk_samples:
+        chunk = np.pad(chunk, (0, chunk_samples - len(chunk)))
+
+    with torch.no_grad():
+        t = torch.from_numpy(chunk).float().unsqueeze(0)
+        ai_prob = float(torch.sigmoid(model(t)).cpu().numpy().reshape(-1)[0])
+
+    return {
+        "ai_prob": round(ai_prob, 4),
+        "duration_s": round(len(y) / sr, 1),
+        "model": MODEL_ID,
+    }