birajsubedi/NepFinSpeech

NepFinSpeech-403: A Domain-Specific Nepali Financial Speech Dataset

Overview

NepFinSpeech-403 is a transcribed speech dataset of 403 Nepali financial voice commands, built as part of the SpeakPay research project — a voice-first digital wallet designed for visually impaired individuals in Nepal.

Existing Nepali ASR resources (OpenSLR, Common Voice) cover general-domain speech but contain very few financial utterances. Financial commands have distinct characteristics — dense numeral sequences (transfer amounts, balances), proper nouns (bank names, recipient names), and a narrow set of recurring sentence structures — that general-purpose models handle poorly. This dataset fills that gap.

Dataset Statistics

Intent Distribution

Additional Statistics

• —237 unique Devanagari numerals (amounts ranging from single digits to six-digit values)
• —Audio format: WAV, mono, 16kHz sample rate
• —Language: Nepali (Devanagari script)
• —Transcription: Manually verified against recorded audio

Usage

from datasets import load_dataset

ds = load_dataset("birajsubedi/NepFinSpeech")

# Access a training sample
sample = ds["train"][0]
print(sample["sentence"])  # Nepali transcript in Devanagari
# sample["audio"] contains the audio array and sampling rate

Collection Methodology

Audio was collected through a purpose-built web data-collection platform. Contributors recorded spoken Nepali financial commands guided by written prompts covering three operation types:

1. 1.Fund transfers — recipient name, amount, and optionally a financial institution
2. 2.Wallet load/deposit — amount to load into a digital wallet
3. 3.Balance enquiries — checking current account balance

Transcripts were manually verified against the recorded audio by the dataset author. Note: transcripts were not independently double-annotated, so no inter-annotator agreement statistic is available. This is a known limitation documented in the accompanying paper.

Benchmark Results

We fine-tuned Whisper large-v2 using LoRA on this dataset. Results on the 60-utterance held-out test set:

• —67.5% relative WER reduction
• —Improvement on 59 out of 60 test utterances (sign test, p = 1.7×10⁻¹²)

The fine-tuned model is available at: birajsubedi/whisper-large-v2-nepali-financial

Intended Use

• —Benchmarking Nepali ASR systems on financial/transactional speech
• —Domain adaptation experiments for low-resource ASR
• —Research on accessible financial technology

Limitations

• —Scale: 403 utterances is small relative to standard ASR corpora. Results should be interpreted as a feasibility demonstration, not a claim of state-of-the-art performance.
• —Speaker diversity: The dataset was collected from a limited pool of contributors and may not fully represent the phonetic variation across Nepali dialects.
• —Single annotator: Transcriptions were verified by a single annotator without independent double-annotation.
• —Intent taxonomy: The "other" category contains a recognizable "funds received" pattern not covered by the three primary intent classes. See the paper for detailed error analysis.

Reproducibility

The full extraction and build pipeline is available in the source repository:

# From the original transcription spreadsheet:
python data/extract_xlsb.py audio-edit1.xlsb -o data/dataset_pairs.json
python data/build_dataset.py

# If you only have this repo (no .xlsb):
# nepfinspeech_dataset.json is already included — use it directly.

License

This dataset is released under CC-BY 4.0.

Citation

@misc{nepfinspeech2025,
  title   = {NepFinSpeech: A Domain-Specific Nepali Financial Speech Dataset},
  author  = {Biraj Subedi},
  year    = {2025},
  url     = {https://huggingface.co/datasets/birajsubedi/NepFinSpeech},
  license = {CC-BY-4.0}
}

Paper

Full technical report with statistical analysis, per-intent breakdown, and error analysis:

• —SpeakPay: Domain-Adaptive LoRA Fine-Tuning of Whisper for Low-Resource Nepali Financial Speech Recognition (LaTeX source)

Contact

Biraj Subedi — Advanced College of Engineering and Management, Tribhuvan University

• —GitHub: github.com/subedibiraj/speakpay
• —Email: contact@biraj-subedi.com.np