AAdonis commited on
Commit
96e7ff9
·
verified ·
1 Parent(s): d82f980

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +1 -29
README.md CHANGED
@@ -774,7 +774,7 @@ Every row consists of a 24 kHz speech utterance paired with a verified human tra
774
 
775
  ## Dataset Overview
776
 
777
- | | |
778
  |---|---|
779
  | Total hours (incl. metadata-only) | 53,635 |
780
  | Embedded-audio hours | ~49,694 |
@@ -854,34 +854,6 @@ WorldSpeech combines MLS-scale hours with FLEURS-scale language coverage and per
854
 
855
  ---
856
 
857
- ## Recommended use cases
858
-
859
- - **Multilingual ASR pretraining and fine-tuning**, especially for languages outside MLS / VoxPopuli coverage.
860
- - **Multilingual TTS pretraining**, after filtering on DNSMOS / SNR / CER.
861
- - **Long-form / streaming ASR**, since `source_start_s` / `source_end_s` allow reconstruction of original session context.
862
- - **Speech enhancement**, using the wide DNSMOS distribution as an in-distribution quality range.
863
-
864
- ---
865
-
866
- ## Known limitations
867
-
868
- - **Domain skew.** Parliamentary, broadcast, and read-audiobook content dominate. Conversational speech is under-represented.
869
- - **Speaker imbalance.** Several configs are dominated by a small number of frequent speakers. No per-speaker labels are provided.
870
- - **Alignment uncertainty.** Human transcripts are aligned via ASR; the `cer` field gives a per-segment quality estimate.
871
- - **Variable audio quality.** Weighted avg DNSMOS OVR is 2.82, with substantial variance. Filter aggressively for TTS-grade work.
872
- - **11 languages have < 10 hours** (`grc`, `mi`, `ml`, `mr`, `nr`, `nso`, `pa`, `ss`, `st`, `ts`, `ve`). Treat these as evaluation signal rather than training data.
873
- - **Five RFA configs are metadata-only.** See embedding caveat above.
874
-
875
- ---
876
-
877
- ## Ethical considerations
878
-
879
- Sources are public records (parliaments, courts, broadcasters, public-domain audiobooks). Speakers are public figures speaking in their official capacity. No private conversations or non-consensual material are included. Configs with non-commercial-only upstream terms (UN, ICC, Vatican Radio, RFA, RFE/RL) are flagged in the Licenses table.
880
-
881
- If you identify content that should not have been included, open an issue and we will action takedowns within 30 days.
882
-
883
- ---
884
-
885
  ## Configs
886
 
887
  113 configs total, sorted by hours descending.
 
774
 
775
  ## Dataset Overview
776
 
777
+
778
  |---|---|
779
  | Total hours (incl. metadata-only) | 53,635 |
780
  | Embedded-audio hours | ~49,694 |
 
854
 
855
  ---
856
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
857
  ## Configs
858
 
859
  113 configs total, sorted by hours descending.