Upload split.py with huggingface_hub

split.py

@@ -0,0 +1,83 @@
+import os
+from collections import Counter
+
+import numpy as np
+import soundfile as sf
+from datasets import load_dataset
+
+dataset = load_dataset(
+    "parquet",
+    data_files={
+        "train": [
+            "data/test-00000-of-00017.parquet",
+            "data/test-00001-of-00017.parquet",
+            "data/test-00002-of-00017.parquet",
+            "data/test-00003-of-00017.parquet",
+            "data/test-00004-of-00017.parquet",
+            "data/test-00005-of-00017.parquet",
+            "data/test-00006-of-00017.parquet",
+            "data/test-00007-of-00017.parquet",
+            "data/test-00008-of-00017.parquet",
+            "data/test-00009-of-00017.parquet",
+            "data/test-00010-of-00017.parquet",
+            "data/test-00011-of-00017.parquet",
+            "data/test-00012-of-00017.parquet",
+            "data/test-00013-of-00017.parquet",
+            "data/test-00014-of-00017.parquet",
+            "data/test-00015-of-00017.parquet",
+            "data/test-00016-of-00017.parquet",
+        ]
+    }
+)
+
+MAX_DURATION = 60.0
+MAX_SPEAKERS = 4
+
+
+def crop_longest_60s(example):
+    starts = example["timestamps_start"]
+    ends = example["timestamps_end"]
+    speakers = example["speakers"]
+    n = len(starts)
+
+    best_left, best_right, best_dur = 0, 0, 0
+    left = 0
+    spk_count = Counter()
+    for right in range(n):
+        spk_count[speakers[right]] += 1
+        while left <= right and (ends[right] - starts[left] > MAX_DURATION or len(spk_count) > MAX_SPEAKERS):
+            spk_count[speakers[left]] -= 1
+            if spk_count[speakers[left]] == 0:
+                del spk_count[speakers[left]]
+            left += 1
+        if left > right:
+            continue
+        dur = ends[right] - starts[left]
+        if dur > best_dur:
+            best_dur = dur
+            best_left, best_right = left, right
+
+    sr = example["audio"]["sampling_rate"]
+    start_sample = int(starts[best_left] * sr)
+    end_sample = int(ends[best_right] * sr)
+    offset = starts[best_left]
+
+    example["audio"]["array"] = example["audio"]["array"][start_sample:end_sample]
+    example["timestamps_start"] = [t - offset for t in starts[best_left:best_right + 1]]
+    example["timestamps_end"] = [t - offset for t in ends[best_left:best_right + 1]]
+    example["speakers"] = example["speakers"][best_left:best_right + 1]
+    return example
+
+
+filenames = [os.path.basename(example["audio"]["path"]) for example in dataset["train"]]
+
+dataset["train"] = dataset["train"].map(crop_longest_60s)
+
+os.makedirs("wavs", exist_ok=True)
+for i, example in enumerate(dataset["train"]):
+    audio = example["audio"]
+    filename = filenames[i]
+    duration = len(audio["array"]) / audio["sampling_rate"]
+    sf.write(f"wavs/{filename}", np.array(audio["array"]), audio["sampling_rate"])
+    num_speakers = len(set(example["speakers"]))
+    print(f"[{i + 1}/16] Saved wavs/{filename}, duration: {duration:.2f}s, segments: {len(example['timestamps_start'])}, speakers: {num_speakers}")