Document DSN hours gap (long-form clips excluded by 1-30s filter)

README.md

@@ -30,6 +30,14 @@ empty/garbage transcripts and undecodable audio were dropped at ingestion. Text
 lightly normalized (Unicode NFC, whitespace/punctuation cleanup) but **tones, digits,
 and code-switching are preserved** — nothing is stripped or transliterated.
 
+**Note on DSN's hours specifically:** DSN African Voices' own published totals per
+language run higher than what's reflected in this source's contribution here. This is
+an intentional effect of the 1–30 second clip-length filter, not an error — DSN's raw
+export includes long-form spontaneous recordings up to several minutes long, dropped
+entirely rather than split (confirmed directly on the Hausa sibling release; see
+[Professor/hausa-speech-data](https://huggingface.co/datasets/Professor/hausa-speech-data)'s
+card for the full investigation). The same DSN ingest script and filter were used here.
+
 ## Format
 
 The dataset ships as **WebDataset-style tar shards** (`shards/shard-00000.tar` …