You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ErfanRou/callcc-new-data-335h

Persian call-centre ASR training windows with Soniox labels, in the same 17-column schema as ErfanRou/callcc-2k-hours-train-soniox: a drop-in for stage_callcc2k.py / build_manifests.py. Machine-labelled, not ground truth.

  • Source: markmuller/call-center-prod-data, newest calls first (by upload commit), excluding every call already in callcc-2k-hours-train-soniox (incl. transcribed calls without rows), callcc-test-1k-samples-soniox and callcc-150h-soniox; dedup key call_id = sha256 of the source wav.
  • Labels: Soniox stt-async-v5, language_hints=["fa"], diarization off, one job per call over a VAD-packed stream: Silero VAD v6 per channel (threshold 0.2, pad 300 ms, pauses up to 3.5 s kept), the speech of both channels laid end to end in time order with 0.25 s of silence between segments, and every token mapped back to its channel and original time. Channel attribution is exact (no diarization); label_source = soniox:stt-async-v5.
  • Windows: the unmodified callcc-2k packer (5-28 s, ~15 s target, natural pauses <= 3.5 s kept, cut in silence), audio cut from the ORIGINAL channel and stored as 16 kHz mono FLAC. A window never contains audio that was not sent to Soniox: stretches that were not sent and had any VAD activity are packer barriers.
  • text - Soniox words normalised (digits->words, Persian ی/ک, no punctuation, ZWNJ kept); text_raw - as emitted; conf_mean/conf_min - Soniox token confidence; text_whisper/wer_whisper/bucket - the production Whisper transcript shipped with the pool, exactly as in callcc-2k (agree < 0.10, informative 0.10-0.50 with conf >= 0.85, contested otherwise, unpaired without Whisper text).
  • tokens/ - raw Soniox tokens per shard: channels (original time, callcc-2k format) plus vadpack (segment map and packed-time tokens), so windows can be re-derived without paying again. meta/calls_meta.parquet - call-level metadata joined on call_id: contains personal data from the CRM (client_metadata_json etc.; retention approved by the data owner 2026-09-17, same as callcc-2k) - keep this repo gated.

Statistics are filled in by callcc_vadpack.py finalize.

Downloads last month
32