Persian call-centre ASR silver set (2,000 audio-hours, 2,288 h shipped): the ErfanRou/callcc-ft150-soniox schema plus two auxiliary columns. Machine-labelled, not ground truth. Built by callcc-silver-5k/soniox_pipeline.py from markmuller/call-center-prod-data: 8 kHz stereo telephony split client-side into agent = channel 0 and customer = channel 1, each channel transcribed separately by Soniox stt-async-v5, words packed into 5-28 s windows (15 s mean, natural pauses kept, short backchannels kept and padded with their own line silence), cut in silence and stored as 16 kHz mono FLAC.

text — training target: Soniox words of the window, normalised (digits→words, Persian ی/ک, no punctuation, ZWNJ kept). text_raw — the same words as Soniox emitted them. conf_mean/conf_min — Soniox token confidence.
text_whisper — auxiliary: the in-house Whisper transcript (DIGI-brain-ASR pre-computed segments shipped with the pool) whose segments fall inside the window, normalised the same way; null when no segment overlaps. Its segment boundaries are approximate (words can be glued at boundaries), so treat it as a second opinion, not a target.
wer_whisper — WER of text_whisper against text (Soniox as reference); null when text_whisper is null.
bucket — agree (< 0.10), informative (0.10-0.50 and conf_mean ≥ 0.85), contested (otherwise), unpaired (no Whisper text). All rows are shipped; filter on bucket/wer_whisper downstream as you see fit.
tokens/ — raw Soniox tokens per shard (gzipped jsonl), meta/calls_meta.parquet — call-level CRM metadata (join on call_id).

Downloads last month: 41

Total file size:

101 GB