anonymous-nsc-author/Neapolitan-Spoken-Corpus

Neapolitan Spoken Corpus (NSC)

A corpus of read Neapolitan speech for ASR evaluation, with a validated Neapolitan–Italian lexicon, LOSO fine-tuning splits, trained LoRA adapters, metric implementations, per-clip results, and error annotations.

This release supersedes the earlier 141-clip single-speaker version of this repository. The earlier release corresponds to Speaker S1 of the present corpus; the old audioData/ and transcripts.csv are replaced by data/audio/ and data/metadata.csv.

Corpus

• 591 clips, 44.2 minutes total
• 4 speakers (S1–S4): 141 / 150 / 150 / 150 clips
• 3 text domains: Blog (147), Play (246), Poetry (198)
• Audio: WAV, 16 kHz, mono, 16-bit PCM. Source recordings were captured at 44.1/48 kHz (WAV and M4A) and converted to 16 kHz WAV for this release.
• Speakers: three male, one female; three adults aged 25–60 and one older adult (60+); all native Neapolitan speakers raised in Campania.
• All texts are read speech from public Neapolitan-language sources (blog prose, theatrical plays, poetry).

Directory map

data/
  audio/                   591 WAV clips (16 kHz mono)
  metadata.csv             clip_id, filename, speaker, domain, duration_s,
                           n_tokens, reference_text
lexicon/
  lexicon_115.csv          validated 115-entry Neapolitan-Italian lexicon
  candidates_118.csv       pre-validation candidate list (118 entries)
  review_log.csv           native-speaker validation log (3 rejected,
                           6 amended, 1 variant-flagged)
  equivalence_classes.json orthographic normalization rules
splits/
  loso_fold{1,2,3,4}/      train.txt / val.txt / test.txt clip-ID lists
adapters/
  C{6,7,8,9}/fold{1,2,3,4}/  LoRA adapters (PEFT format), one per
                             condition x fold
metrics/
  ier.py, ier_alignment.py, normalize.py    IER implementation
  glotlid_ratio.py                          GlotLID-ratio metric
  scoring_pipeline.py                       WER/CER + end-to-end scoring
  reproduce.py                              reproduces headline numbers
results/
  per_clip_uncapped.csv    uncapped per-clip WER/CER, all conditions
  zero_shot/               per-clip results, conditions C1-C5
  lora/                    per-clip results, conditions C6-C9 (4 folds each)
annotations/
  error_analysis_pass1.csv 120-clip error annotations, first pass
  pass2/                   blind second-pass annotations, taxonomy card,
                           blind key, taxonomy-differences README
rebuttal/
  R1/                      zero-shot baselines for two non-Whisper systems
                           (SeamlessM4T-v2-large, MMS-1b-all)
  R3/                      lexicon-subsampling sensitivity analysis

Evaluation protocol (LOSO)

Fine-tuning conditions C6–C9 use leave-one-speaker-out cross-validation: each fold holds out one speaker's clips as the test set; the remaining three speakers' clips are split 90/10 into train/validation (seed 42). Fold definitions are in splits/. Reported corpus-level values pool the four held-out test sets, which together cover all 591 clips exactly once.

Reproducing the paper's numbers

cd metrics
python reproduce.py

This recomputes the headline corpus values (C1 WER 0.8124, C1 IER 0.1375, C9 IER 0.0612) from the raw per-clip references/hypotheses in results/ using only the released metric code and lexicon. See metrics/README.md for metric definitions and the full mapping to the results tables.

Python 3.10+ is required. requirements.txt pins the full environment used for the paper's experiments; reproduce.py itself needs only the standard library. The pinned torch==2.11.0+cu128 build installs from the PyTorch CUDA 12.8 index.

Adapters

Each adapters/C*/fold*/ directory is a PEFT LoRA adapter (adapter_config.json + adapter_model.safetensors; r=16, alpha=32, q_proj/v_proj). C6–C8 adapt openai/whisper-small; C9 adapts openai/whisper-medium. Load with:

from peft import PeftModel
from transformers import WhisperForConditionalGeneration

base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
model = PeftModel.from_pretrained(base, "adapters/C6/fold1")

Dataset viewer

The viewer is configured to display data/metadata.csv (one row per clip). Audio files are under data/audio/ and are named by the filename column.

License and consent

This dataset is released under a Creative Commons Attribution-NonCommercial (CC BY-NC 4.0) license. All participants provided informed consent, and no sensitive or personal information is included. See LICENSE.md.