Full-channel benchmark set for Persian call-centre ASR: one row per channel of a call (0 = agent, 1 = customer), with the complete 16 kHz mono channel audio and the complete Soniox stt-async-v5 transcript of that channel, rebuilt from the raw tokens of ErfanRou/callcc-test-1k-windowed (no re-transcription). Use it to evaluate the serving path (whole-channel input, model-side VAD/chunking) with corpus-level WER/CER — see eval_full_call.py in the kit.

text — normalised reference (digits→words, Persian ی/ک, no punctuation, ZWNJ kept); text_raw — verbatim Soniox.
words — Soniox word timestamps + confidence (tokens over 4 s are capped at 1.5 s; the rest was silence/hold).
text_whisper — verbatim production Whisper output for the same channel (DIGI-brain-ASR = lora_merged-ct2, beam 5, VAD on, server-side normalisation on, no LLM polish), taken from markmuller/call-center-prod-data transcripts/<xx>/<call_id>.json → channels[channel].text (source revision 87b8adeb); null where the model emitted no segments (22 rows). It is NOT passed through the reference normaliser — fold both sides at scoring time.
crm_metadata — the call's CRM/client metadata as JSON (retention approved by the data owner), null if the call has none. The reference is machine-generated: treat absolute WER as agreement with Soniox; use for relative comparisons and regressions.

Downloads last month: 102

Total file size:

2.38 GB