RapBank — Final Dataset

Processed and filtered subset of the RapBank corpus, ready for model training.

Data Source

Raw audio and transcripts originate from:

./work/corpus/rap/raw/

Audio files are source-separated vocal clips segmented by VAD, stored as .flac under vocal_cut/{video_id}/. Transcripts were produced by Whisper and cross-validated against Parakeet TDT; clips longer than 30 s were removed. The full preprocessing pipeline is documented in ./work/corpus/rap/raw/README.md.

Dataset Format

Each split (train/, valid/, test/) contains two manifest files:

`utt2fpath.tsv`

Maps utterance ID to its audio file path (relative to ./work/corpus/rap/raw/):

{utt_id}\t{relfpath}

Example:

TTEiUDYwk84_1_29390    vocal_cut/TTEiUDYwk84/TTEiUDYwk84_1_29390.flac

`utt2asrtext.tsv`

Maps utterance ID to its ASR transcript (Whisper, cross-validated against Parakeet):

{utt_id}\t{transcript}

Example:

TTEiUDYwk84_1_29390    Who want it, nigga? We the violentest...

The utt_id is the filename stem of the audio clip and is consistent across all utt2*.tsv files within a split.

Planned: Arrow Files

.arrow (Apache Arrow / HuggingFace Datasets format) files will be generated from the utt2*.tsv manifests in a future preparation step. Each shard will bundle the raw audio waveform alongside all metadata columns.

Current Contents

dataset_final/
├── train/
│   ├── utt2fpath.tsv
│   └── utt2asrtext.tsv
├── valid/
│   ├── utt2fpath.tsv
│   └── utt2asrtext.tsv
└── test/
    ├── utt2fpath.tsv
    └── utt2asrtext.tsv

No .arrow files have been generated yet.

Generating the Splits

python egs/taste_s/rslm/scripts/data/prepare_data.py \
    --src_tsv work/corpus/rap/raw/final_valid_ids.tsv \
    --tgt_dir work/corpus/rap/dataset_final \
    --valid_size 1000 \
    --test_size  1000