DimitarV/bulgarian-parliament-moss-1to1

Bulgarian Parliament MOSS 1-to-1

Private machine-labelled training candidates under construction. Not gold data or a held-out evaluation set. Public redistribution terms remain unverified.

Source: DimitarV/eurospeech-bg-single-speaker at 6aa43432891867bcd249c0fe944ac6286888231b, derived from disco-eth/EuroSpeech. Reference transcripts are parliamentary stenographic text. Speaker identities are inferred clusters, not named or human-verified speakers.

Active batch list

Use catalog.json to select batches. Do not train on every tar in this repo. It gives the currently active verified archives, checksums, revisions, and actual original/mixed hours. pilot-000 is excluded. batch-001-v2 supersedes batch-001; the older archive remains for reproducibility and must not be counted or loaded alongside its replacement.

The current run screens approximately 40 candidate hours per batch, retaining all qualified originals, with approximately equal synthetic-dialogue hours (duration balance tolerance 0.5%). There is no 25-hour retention cap. The 1,000 original + 1,000 mixed hours discussed for this project is a stretch target, not a statement of uploaded size. Source mixtures do not create new independent recorded speech.

Layout

Extract each batches/BATCH/BATCH.tar into its own batch directory:

• train.jsonl: MOSS conversation-format targets and relative audio paths.
• records.jsonl: timed speaker spans and source provenance.
• audio/source/: source FLACs used in this batch.
• audio/dialogue/: synthetic mixed FLACs.
• quality.jsonl: per-source filtering and alignment evidence.
• summary.json: actual hours and structural validation.

Do not flatten batch directories. Row IDs include batch prefixes.

Current quality policy: expansion-v2-timing-flag

• Original and fresh-ASR normalized CER must both be below 10%. Existing CER tiers <=2%, 2–5% and 5–10% remain in the metadata. The reference is retained.
• Fresh full-clip Whisper large-v3 uses batches of eight, beam size five, and serial fallback on uncertain results; source boundaries remain independent.
• pyannote community-1 rejects detected secondary speech over 300 ms and clips with under 500 ms primary speech. These remain model estimates.
• The reference alignment must contain every word with valid ordered timing and direct reliable outer boundaries. Invalid alignments are screened first.
• Whisper-vs-alignment outer boundary disagreement is a diagnostic flag only. It is not evidence by itself of missing speech. A 19-clip/five-minute human listening audit found the sampled disagreements compatible with silence boundary differences and no unaccounted speech. This small targeted audit is not a corpus-wide error estimate. Old batch-001 used a stricter rejection policy and is superseded.
• Mixtures retain simultaneous speaker spans, source provenance, activity-based overlap estimates and the existing sentence-start constraints on holds.

Training sources may overlap earlier project training runs. Do not use these batches as a held-out benchmark. Cross-clip speaker cluster identities and machine timing should be audited before constructing an evaluation set.

Storage and publication

Each archive is uploaded privately, then read back at the exact Hub commit and verified by SHA-256 before generated local data is reclaimed. Original cached project sources and model checkpoints are preserved; newly downloaded source caches are recyclable from the pinned upstream dataset.

No permissive redistribution licence is asserted. The upstream Bulgarian licensing entry points to a presidential website, which does not establish permission for these parliamentary recordings. Source terms need verification before any public release.