Neapolitan-Spoken-Corpus (NSC)

Neapolitan-Spoken-Corpus (NSC) is the first publicly available speech corpus designed specifically for benchmarking Automatic Speech Recognition (ASR) systems on Neapolitan, a low-resource Romance dialect of Southern Italy. It includes 141 sentence-level audio recordings along with gold-standard orthographic transcriptions.

The dataset was created to address the lack of computational resources for dialectological research and the development of equitable speech technologies.

Dataset Structure

Neapolitan-Spoken-Corpus/
├── audioData/
│   ├── 002.m4a
│   ├── 003.m4a
│   ├── ...
│   └── 142.m4a
├── code/
│   ├── generate_json.py
│   ├── transcribe_whisper.py
│   └── evaluate_metrics.py
├── .gitattributes
├── README.md
├── requirements.txt
└── transcripts.csv

Intended Uses & Limitations

The dataset is primarily intended for evaluating and developing ASR systems that support dialectal languages, particularly those with minimal computational resources. It provides a benchmark for dialect-aware speech recognition and can also support linguistic research in computational dialectology and language preservation.

How to Use

To use this dataset and its associated scripts:

# Clone repository
git clone https://huggingface.co/datasets/anonymous-nsc-author/neapolitan-spoken-corpus
cd neapolitan-spoken-corpus
# Install dependencies
pip install -r requirements.txt
# (Optional) Generate sentences.json
python code/generate_json.py
# Transcribe audio files with Whisper ASR (requires OPENAI_API_KEY)
export OPENAI_API_KEY=your-key-here
python code/transcribe_whisper.py
# Evaluate transcription accuracy metrics (WER, BLEU, etc.)
python code/evaluate_metrics.py

Evaluation Results

The dataset was evaluated using OpenAI's Whisper model with the language set to Standard Italian. The results indicate significant performance degradation on Neapolitan dialect speech:

Metric	Mean	Std Dev	Min	Max
WER (1 - WER similarity)	0.1306	0.1654	0.0000	0.9091
Levenshtein (normalized)	0.6360	0.1375	0.0870	0.9804
BLEU	0.0436	0.0961	0.0000	0.8932
Jaccard	0.1078	0.1294	0.0000	0.8333

Ethical Considerations

All participants involved in creating this dataset provided explicit informed consent. Audio and transcription data include no sensitive, private, or personally identifiable information.