Merge branch 'jsalt_2023' of https://github.com/revdotcom/speech-data…

…sets into jsalt_2023

longform_reconstitution/README.md

@@ -1,13 +1,22 @@
 # JSALT 2023 Long-Form ASR Corpora
 
-Long-form ASR corpora created by reconstituting transcriptions from existing, open ASR corpora. All speech and transcriptions are in English.
+Long-form ASR corpora created by reconstituting transcriptions from existing, open ASR corpora. All transcriptions are in English. This repository contains only time-aligned transcriptions - audio can be downloaded from the following locations:
 
 1. GigaSpeech: https://github.com/SpeechColab/GigaSpeech/
 2. TED-LIUM: https://www.openslr.org/7/
 3. VoxPopuli: https://github.com/facebookresearch/voxpopuli
 
+All data in this repository made available as Lhotse supervisions - see https://github.com/lhotse-speech/lhotse/tree/master for more information. 
+
 ## GigaSpeech
 
+Currently available: Lhotse supervisions for four versions of the GigaSpeech M subset. 
+
+1. `gigaspeech/gigaspeech_supervisions_m.jsonl.gz` - original segmentation of the transcriptions of the GigaSpeech M subset (Note: not time-aligned). Average segment length 4s 
+2. `gigaspeech/gigaspeech_supervisions_m_lf.jsonl.gz` - full reconstitutions of the transcriptions of the GigaSpeech M subset. Average segment length 10s
+3. `gigaspeech/gigaspeech_supervisions_m_lf_15.jsonl.gz` - full reconstitution, followed by re-segmentation with a maximum segment length of 15s. Average segment length 7s
+4. `gigaspeech/gigaspeech_supervisions_m_lf_30.jsonl.gz` - full reconstitution, followed by re-segmentation with a maximum segment length of 30s. Average segment length 9s
+
 ## TED-LIUM
 
-## VoxPopuli
+## VoxPopuli