theusamaaslam/urdu-callsqa-private

Calls QA Urdu Speech

Internal human call speech, curated separately from the organic Urdu collection. Contains 63,050 clips / 391.1683 hours. Earlier strict clips measure 1-10 seconds; broader candidates measure 1-300 seconds. Access must remain private or gated. The exporter refuses an ungated public destination and never changes repository visibility or approval settings. For a public gated repository, the dataset card is public while dataset files require access through the repository's configured gating policy. Recordings may contain personal details and are authorized for internal use only; no public redistribution grant is implied.

Schema: id, audio, text, speaker_id, source, duration_seconds, language, is_synthetic. Audio bytes are embedded as lossless FLAC in Parquet; source sample rates are preserved. text preserves the existing Roman Urdu transcription and English code switching without transliteration or rewriting. source=callsqa_human, language=ur, is_synthetic=false.

The collector reads completed Calls QA records and bounded WAV ranges without changing the production application, database, services, or dependencies. Only complete existing HUMAN transcript spans are used. The owner authorized broader candidate collection on 2026-11-08 after reporting manual transcript/timestamp spot checks. Consecutive turns are grouped toward 30 seconds; complete longer turns up to 300 seconds are retained without assigning guessed text to cuts. Tied timestamps are grouped, and conflicting timestamp neighborhoods are excluded instead of rejecting the entire call. Known IVR labels, timed IVR overlap and substantial matching IVR text are excluded. With untimed IVR, the existing HUMAN field's separation is trusted and explicitly flagged; this does not certify that all residual IVR has been removed. Missing/failed source-quality flags, clipping, silence and voiced boundaries are audit warnings rather than blanket rejections. Corrupt, silent, non-finite and invalid-range audio remains excluded. No new model calls are made. Spoken words are preserved; speaker labels/annotations are removed and consecutive references may be joined with newlines. Original source audio is not archived and the production retention schedule remains unchanged. Timestamp and energy checks do not establish word-level alignment or human gold transcription; a sampled alignment audit is required before a training release.

speaker_id is a pseudonymous call-and-role label, not a verified global speaker identity. The same agent may have different IDs across calls. Cluster speakers before constructing speaker-disjoint validation/test splits. Grouped conversations can contain multiple speakers (role mixed). These broader rows are candidates for later training preparation, not independently verified single-speaker TTS clips. Their policy and quality flags are in the provenance.

All additions are train-only. The original organic dataset and its locked validation/test splits are untouched. Future merges require deduplication, speaker-overlap review, and an explicit release decision. Provenance and automated-check results are recorded under metadata/callsqa/.