From 3ca3fed561c20f60d3e4594ccc874c3af87dc4e8 Mon Sep 17 00:00:00 2001
From: Ronald Tse <ronald.tse@ribose.com>
Date: Thu, 8 Oct 2026 15:58:55 +0800
Subject: [PATCH] =?UTF-8?q?feat:=20WO09=20stage=200=20=E2=80=94=20FLEURS?=
 =?UTF-8?q?=20he=5Fil=20audio=20sample=20fetch=20(CC-BY-4.0)=20to=20unbloc?=
 =?UTF-8?q?k=20the=20parked=20probe?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 fetch_hebrew_audio_hf.py | 47 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 47 insertions(+)
 create mode 100644 fetch_hebrew_audio_hf.py

diff --git a/fetch_hebrew_audio_hf.py b/fetch_hebrew_audio_hf.py
new file mode 100644
index 0000000..1c6e22d
--- /dev/null
+++ b/fetch_hebrew_audio_hf.py
@@ -0,0 +1,47 @@
+# /// script
+# requires-python = ">=3.11"
+# dependencies = [
+#   "datasets>=3.0",
+#   "soundfile>=0.12",
+#   "numpy<2",
+# ]
+# ///
+"""WO09 un-park, stage 0: fetch a small Hebrew audio sample.
+
+FLEURS he_il (CC-BY-4.0, Google/CMU/Microsoft), train split, streamed —
+20 clips saved as wav into the bucket for the WO09 probe. The probe
+itself still fetches nothing (WO09 contract).
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import soundfile as sf
+
+OUT = Path("/ckpt/audio-sample")
+N = 20
+
+
+def main() -> None:
+    OUT.mkdir(parents=True, exist_ok=True)
+    from datasets import load_dataset
+
+    ds = load_dataset("google/fleurs", "he_il", split="train", streaming=True)
+    saved = 0
+    for row in ds:
+        if saved >= N:
+            break
+        audio = row["audio"]
+        if len(audio["array"]) < 16000:  # skip <1s clips
+            continue
+        p = OUT / f"fleurs-he-{saved:03d}.wav"
+        sf.write(str(p), audio["array"], audio["sampling_rate"])
+        saved += 1
+        if saved % 5 == 0:
+            print(f"[fetch] {saved}/{N}", flush=True)
+    print(f"[done] {saved} clips -> {OUT}", flush=True)
+
+
+if __name__ == "__main__":
+    main()