This is the most valuable corpus in the collection and the smallest. It is the only one that is both human-transcribed and spontaneous telephony -- the register Chinar-F8 actually targets. Adding 134.7 h of it to the training mix moved WER on human-transcribed calls from 43.39% to 35.08%, the largest single gain any data change has produced here.

Fields

id, audio (FLAC), text, text_normalised, seconds, source, label_type, text_group, speaker_id.

Splits are component-safe: clips sharing a transcript, a speaker or a source recording land on the same side, so a model cannot see a paraphrase of its test set during training.

Limitations

Transcribers write clean sentences and drop fillers, so the text is accurate but not verbatim. Russian and code-switched turns were excluded. Real calls contain personal data; this dataset is private and is not to be redistributed.

Labelling

field	value
`label_origin`	`human-transcript`
`speech_register`	`spontaneous`
`channel`	`telephony-8k`
`provenance`	`documented`

People listened to the audio and wrote what they heard. The process is recorded rather than assumed: 4,301 calls carry named validators and weekly batch assignments.

_{label_origin distinguishes text that existed before the audio (script,
exact by construction) from text written by a listener (human-transcript, high
but edited) from machine output (asr:<vendor>, bounded by that vendor's error
rate).}

Downloads last month: 19

Total file size:

5.86 GB