calls-train-human

ASR training clips from KejueAI/human-call-recordings (internal client call recordings, private). 654.0 h kept.

Human-agent call-centre calls (mono 8 kHz, upsampled to 16 kHz). Both sides are real people, so both are kept.

Labels = Soniox stt-async-v5 transcripts (clean verbatim). Clip bounds = Soniox speaker-turn times + 0.15 s pad, trimmed to Silero-VAD speech + 0.25 s. No CTC aligner (avoids the digit-damage bug of the old pipeline). Turns > 25 s were dropped (no word timestamps to split them).

Filters: bot side removed; no detectable speech; speaking rate outside 2-30 chars/s; duration outside 0.5-30 s. Dropped: no_speech_h 11.6 h, speaking_rate_h 2.1 h.

accent column: English speech -> CommonAccent ECAPA vote per (call, speaker) when >= 60 % agree, else english_unknown; a speaker who also speaks Arabic in the same call -> arabic (the accent model has no Arabic class). Arabic speech -> call dialect from manifest_v2 (Gemma text classification). Other scripts -> other_language.

Not a benchmark: these calls overlap the existing KejueAI/asr-* training repos.

Format: parquet, 500-row row groups (Cohere training format). Columns: audio (16 kHz mono FLAC bytes), id, text, duration, accent, speaker_id (call_speaker_channel), gender (empty), source, lang, reference_text (empty), meta (JSON: client, call_id, start/end, bot_score, accent_votes, call_dialect, ...).

client	accent	hours
JoAcademy	levantine	252.0
SA-	saudi	157.8
SA-	levantine	149.0
March12-eCommerce	levantine	29.1
SA-	arabic	7.7
Jord_Ordering_8k_done	levantine	7.5
March12-eCommerce	iraqi	6.3
JoAcademy	other_language	6.0
March12-eCommerce	saudi	5.2
SA-	iraqi	4.0
SA-	mixed	3.1
JoAcademy	saudi	2.7
SA-	msa	2.7
March12-eCommerce	gulf	2.7
SA-	egyptian	2.5
SA-	gulf	1.8
SA-	other_language	1.7
March12-eCommerce	mixed	1.5
SA-	english_unknown	1.5
SA-	us	1.2
JoAcademy	msa	1.1
SA-	england	1.0
Jord_Ordering_8k_done	iraqi	1.0
JoAcademy	england	0.6
Jord_Ordering_8k_done	saudi	0.5