burkimbia/spg_series01

SPG Series - Moore aligned speech (EP1)

Sentence-level audio <-> Moore text pairs, force-aligned from the bilingual TV show Sid Pa Gilmde (episode 1, "Falez DK"). Built for Moore ASR / TTS.

• Audio: 16 kHz mono wav clips, one per sentence.
• Text: Moore (mos), in spoken order; source_fr is the French reference.
• Alignment: MMS forced alignment (ctc-forced-aligner, uroman), then pause-trimming (densest_run, gap > 1.5 s) to drop mis-anchored words across silences / laughter.

Columns

column	meaning
audio	16 kHz clip
text	Moore sentence (trimmed to the aligned run)
source_fr	French reference
start, end, duration	span in the source video (s)
cps	chars / second (alignment sanity)
keep	passes the gate: 0.5 <= duration <= 30 s and 5 <= cps <= 25
trimmed	True if pause-trimming shortened the sentence
align_score	mean per-word MMS alignment score (higher = more confident)

Use keep == True for training. The keep == False rows are kept for audit.

Provenance & limits

Text from bia-datasets-text (SPG series, PR #40). The text is a cleaned transcription, not strictly verbatim: laughter, fillers and untranscribed speech in the audio cause residual drift, flagged by low align_score / out-of-range cps. Source video: https://www.youtube.com/watch?v=qu6PXQfOy9c

fr_marker column

Sentence-initial French discourse marker detected from source_fr (empty if none). French function words spoken in the audio were normalized to Moore in text (Donc/Maintenant -> Rẽnd, Mais -> La, Parce que -> Bala, Voilà -> Walaa; Ok kept). BIA-Whisper, trained on the same normalized text, does not reproduce the French form, so source_fr is the reliable detector for these 37 rows.