giangndm commited on
Commit
6f5d375
·
verified ·
1 Parent(s): ebc294a

Audio Compress: sample real audio to report sample rate / channels / codec

Browse files
Files changed (1) hide show
  1. features/audio_compress.py +90 -6
features/audio_compress.py CHANGED
@@ -278,9 +278,11 @@ def resolve_source(repo: str, config: str, token: str, revision: Optional[str])
278
  first = next(iter(splits.values()))[0]
279
  fs = HfFileSystem(token=token)
280
  with fs.open(f"datasets/{repo}@{quote(rev, safe='')}/{first}", "rb") as fh:
281
- schema = pq.ParquetFile(fh).schema_arrow
282
- features = features_from_schema(schema)
283
- audio_cols = {name: feat.get("sampling_rate") for name, feat in features.items() if feat.get("_type") == "Audio"}
 
 
284
 
285
  return {
286
  "repo": repo,
@@ -292,9 +294,85 @@ def resolve_source(repo: str, config: str, token: str, revision: Optional[str])
292
  "rows": rows,
293
  "features": features,
294
  "audio_cols": audio_cols,
 
295
  }
296
 
297
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
298
  def features_from_schema(schema: pa.Schema) -> dict:
299
  """Features (as a dict) from the parquet `huggingface` metadata, falling back to the arrow schema."""
300
  raw = (schema.metadata or {}).get(b"huggingface")
@@ -329,8 +407,9 @@ def source_summary_md(meta: dict) -> str:
329
  lines.append(f"| `{split}` | {len(files):,} | {human_size(size) if size else '?'} | {f'{rows:,}' if rows else '?'} |")
330
  lines.append("")
331
  if meta["audio_cols"]:
332
- cols = ", ".join(f"`{c}` ({sr or '?'} Hz)" for c, sr in meta["audio_cols"].items())
333
- lines.append(f"- 🎧 **Audio columns:** {cols}")
 
334
  else:
335
  lines.append("- ❌ **No `Audio` column found** in this subset.")
336
  others = [c for c in meta["features"] if c not in meta["audio_cols"]]
@@ -556,7 +635,12 @@ def load_config(oauth_token: Optional[gr.OAuthToken], session_token: Optional[st
556
  return *empty, f"❌ Cannot read subset `{config}`:\n\n```\n{type(exc).__name__}: {exc}\n```"
557
  splits = list(meta["splits"])
558
  audio = list(meta["audio_cols"])
559
- meta["sampling_rate"] = next((sr for sr in meta["audio_cols"].values() if sr), None)
 
 
 
 
 
560
  return (
561
  gr.update(choices=splits, value=splits[0] if splits else None),
562
  gr.update(choices=audio, value=audio),
 
278
  first = next(iter(splits.values()))[0]
279
  fs = HfFileSystem(token=token)
280
  with fs.open(f"datasets/{repo}@{quote(rev, safe='')}/{first}", "rb") as fh:
281
+ pf = pq.ParquetFile(fh)
282
+ schema = pf.schema_arrow
283
+ features = features_from_schema(schema)
284
+ audio_cols = {name: feat.get("sampling_rate") for name, feat in features.items() if feat.get("_type") == "Audio"}
285
+ probe = {col: probe_audio(pf, col) for col in audio_cols}
286
 
287
  return {
288
  "repo": repo,
 
294
  "rows": rows,
295
  "features": features,
296
  "audio_cols": audio_cols,
297
+ "probe": probe,
298
  }
299
 
300
 
301
+ #: Max compressed size of the column chunk read to sample the audio (keeps Load fast).
302
+ PROBE_MAX_BYTES = 64 * 1024 * 1024
303
+ PROBE_ROWS = 8
304
+
305
+
306
+ def probe_audio(pf: pq.ParquetFile, col: str) -> dict:
307
+ """Read a few audio values of the first row group and inspect them with `soundfile.info`.
308
+
309
+ Many datasets declare `Audio()` without a sampling rate (audio kept at its native rate),
310
+ so the metadata alone cannot tell the sample rate / channels / codec.
311
+ """
312
+ try:
313
+ if pf.metadata.num_row_groups == 0:
314
+ return {"error": "empty shard"}
315
+ rg = pf.metadata.row_group(0)
316
+ idx = pf.schema_arrow.get_field_index(col)
317
+ chunk = sum(
318
+ rg.column(i).total_compressed_size
319
+ for i in range(rg.num_columns)
320
+ if rg.column(i).path_in_schema.split(".")[0] == col
321
+ ) if idx >= 0 else 0
322
+ if chunk > PROBE_MAX_BYTES:
323
+ return {"error": f"first row group too large to sample ({human_size(chunk)})"}
324
+ values = pf.read_row_group(0, columns=[col]).column(col).to_pylist()
325
+ except Exception as exc: # noqa: BLE001
326
+ return {"error": f"{type(exc).__name__}: {exc}"}
327
+
328
+ rates, channels, codecs, durations, errors = set(), set(), set(), [], []
329
+ for value in [v for v in values if v and v.get("bytes")][:PROBE_ROWS]:
330
+ try:
331
+ info = sf.info(io.BytesIO(value["bytes"]))
332
+ except Exception: # noqa: BLE001
333
+ ext = PurePosixPath(value.get("path") or "").suffix or "unknown format"
334
+ errors.append(ext)
335
+ continue
336
+ rates.add(info.samplerate)
337
+ channels.add(info.channels)
338
+ codecs.add(f"{info.format} {info.subtype}")
339
+ durations.append(info.duration)
340
+ if not rates and not errors:
341
+ return {"error": "no embedded audio bytes in the sampled rows"}
342
+ return {
343
+ "rates": sorted(rates),
344
+ "channels": sorted(channels),
345
+ "codecs": sorted(codecs),
346
+ "avg_seconds": sum(durations) / len(durations) if durations else None,
347
+ "sampled": len(durations) + len(errors),
348
+ "undecodable": errors,
349
+ }
350
+
351
+
352
+ def describe_audio(col: str, declared_sr: Optional[int], probe: Optional[dict]) -> str:
353
+ """One line describing an audio column: declared metadata + what the samples contain."""
354
+ declared = f"declared {declared_sr:,} Hz" if declared_sr else "no sampling_rate declared"
355
+ probe = probe or {}
356
+ if probe.get("error"):
357
+ return f"`{col}` — {declared}; could not sample the audio ({probe['error']})"
358
+ parts = []
359
+ if probe.get("rates"):
360
+ parts.append(" / ".join(f"{sr:,} Hz" for sr in probe["rates"]))
361
+ if probe.get("channels"):
362
+ parts.append(" / ".join({1: "mono", 2: "stereo"}.get(c, f"{c} ch") for c in probe["channels"]))
363
+ if probe.get("codecs"):
364
+ parts.append(", ".join(probe["codecs"]))
365
+ if probe.get("avg_seconds"):
366
+ parts.append(f"~{probe['avg_seconds']:.1f} s / clip")
367
+ line = f"`{col}` — {' · '.join(parts)} _(sampled {probe.get('sampled', 0)} rows; {declared})_"
368
+ if probe.get("undecodable"):
369
+ line += (
370
+ f"\n - ⚠️ {len(probe['undecodable'])} sampled clip(s) cannot be decoded by libsndfile "
371
+ f"({', '.join(sorted(set(probe['undecodable'])))}) — the job would fail on them."
372
+ )
373
+ return line
374
+
375
+
376
  def features_from_schema(schema: pa.Schema) -> dict:
377
  """Features (as a dict) from the parquet `huggingface` metadata, falling back to the arrow schema."""
378
  raw = (schema.metadata or {}).get(b"huggingface")
 
407
  lines.append(f"| `{split}` | {len(files):,} | {human_size(size) if size else '?'} | {f'{rows:,}' if rows else '?'} |")
408
  lines.append("")
409
  if meta["audio_cols"]:
410
+ lines.append("- 🎧 **Audio columns:**")
411
+ for col, sr in meta["audio_cols"].items():
412
+ lines.append(f" - {describe_audio(col, sr, meta.get('probe', {}).get(col))}")
413
  else:
414
  lines.append("- ❌ **No `Audio` column found** in this subset.")
415
  others = [c for c in meta["features"] if c not in meta["audio_cols"]]
 
635
  return *empty, f"❌ Cannot read subset `{config}`:\n\n```\n{type(exc).__name__}: {exc}\n```"
636
  splits = list(meta["splits"])
637
  audio = list(meta["audio_cols"])
638
+ # For the estimate: declared rate, else the most common sampled rate / channel count.
639
+ probes = [p for p in meta.get("probe", {}).values() if p.get("rates")]
640
+ meta["sampling_rate"] = next((sr for sr in meta["audio_cols"].values() if sr), None) or (
641
+ max(probes[0]["rates"]) if probes else None
642
+ )
643
+ meta["channels"] = max(probes[0]["channels"]) if probes else None
644
  return (
645
  gr.update(choices=splits, value=splits[0] if splits else None),
646
  gr.update(choices=audio, value=audio),