README.md

metadata

license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
language:
  - en
  - he
tags:
  - speech-to-text
  - stt
  - evaluation
  - technical-vocabulary
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*

Small STT Eval Audio Dataset

A small speech-to-text evaluation dataset containing 92 audio samples with ground truth transcriptions. Designed for evaluating STT systems on technical vocabulary, code-switching (English/Hebrew), and various speaking styles.

Dataset Description

This dataset contains audio recordings with accompanying transcriptions across multiple categories:

Category	Count	Description
tech_github	5	GitHub-related technical vocabulary
tech_huggingface	4	Hugging Face platform terminology
tech_docker	5	Docker and containerization terms
hebrew_daily	10	English with Hebrew words (daily life)
hebrew_food	3	English with Hebrew food terms
ai_ml	9	AI/ML technical vocabulary
local_tools	8	Local development tools
conversational	10	Casual conversational speech
narrative	6	Narrative/storytelling style
instructions	7	Instructional content
tech_linux	6	Linux system administration
tech_api	4	API and web services
tech_python	5	Python programming
mixed_workflow	5	Mixed technical workflows
mixed_locale	2	Mixed locale content
tech_web	2	Web development
tech_data	1	Data processing

Audio Specifications

• Format: WAV (PCM signed 16-bit little-endian)
• Sample Rate: 16kHz
• Channels: Mono
• Average Duration: ~5-10 seconds per sample

Dataset Structure

data/
  ├── metadata.csv
  ├── 001_tech_github.wav
  ├── 002_tech_github.wav
  └── ...

The metadata.csv contains:

• file_name: Audio filename
• transcription: Ground truth transcription
• category: Content category

Usage

from datasets import load_dataset

dataset = load_dataset("danielrosehill/Small-STT-Eval-Audio-Dataset")

# Access a sample
sample = dataset["train"][0]
print(sample["transcription"])
# Play audio: sample["audio"]

Intended Use

This dataset is intended for:

• Evaluating STT model accuracy on technical vocabulary
• Testing code-switching (English/Hebrew) recognition
• Benchmarking STT systems on varied speaking styles
• Development and testing of speech recognition pipelines

Recommended Evaluation Packages

For WER (Word Error Rate) evaluation, we recommend using text normalization to handle variations in number formatting, punctuation, and casing:

• whisper-normalizer: Text normalization for STT evaluation (handles "3000" vs "three thousand", punctuation, casing)
• werpy: WER calculation with detailed error analysis

from whisper_normalizer.english import EnglishTextNormalizer
from werpy import wer

normalizer = EnglishTextNormalizer()

# Normalize both reference and hypothesis before comparison
reference = normalizer(ground_truth)
hypothesis = normalizer(model_output)

error_rate = wer(reference, hypothesis)

Limitations

• Small dataset size (92 samples)
• Single speaker
• Controlled recording environment
• Limited Hebrew vocabulary (loan words only, not full Hebrew speech)

License

CC-BY-4.0