Numb3rs

# Numb3rs - Numbers Speech Benchmark (Dataset) A speech dataset for text normalization (TN) and inverse text normalization (ITN) tasks, containing paired written/spoken forms with corresponding synthetic audio. ## Dataset Creation This dataset was created through the following pipeline: 1. **Source Data**: Text normalization pairs were derived from the [Google Text Normalization dataset](https://www.kaggle.com/datasets/google-nlu/text-normalization), containing written forms (e.g., "$100") and their spoken equivalents (e.g., "one hundred dollars"). 2. **Audio Generation**: Audio was synthesized using [**Magpie TTS**](https://build.nvidia.com/nvidia/magpie-tts-multilingual/modelcard) (NVIDIA's expressive multilingual text-to-speech model), with utterances distributed across **6 predefined voices** to ensure speaker diversity. 3. **Human Verification**: All generated samples were manually verified by human annotators. Only entities that passed quality review were retained in the final dataset. ## Dataset Statistics | Category | Samples | Total Duration | Avg Duration | Description | |----------|---------|----------------|--------------|-------------| | ADDRESS | 885 | 18.7 min | 1.26s | Highway/road identifiers (e.g., "A6" → "a six") | | CARDINAL | 780 | 14.5 min | 1.11s | Cardinal numbers (e.g., "42" → "forty two") | | DATE | 977 | 30.6 min | 1.88s | Date expressions (e.g., "Jan 1, 2020" → "january first twenty twenty") | | DECIMAL | 928 | 24.9 min | 1.61s | Decimal numbers (e.g., "3.14" → "three point one four") | | DIGIT | 771 | 17.8 min | 1.39s | Digit sequences (e.g., "123" → "one two three") | | FRACTION | 884 | 23.4 min | 1.59s | Fractional values (e.g., "1/2" → "one half") | | MEASURE | 914 | 27.7 min | 1.82s | Measurements (e.g., "5 kg" → "five kilograms") | | MONEY | 775 | 26.8 min | 2.07s | Currency amounts (e.g., "$100" → "one hundred dollars") | | ORDINAL | 957 | 14.3 min | 0.90s | Ordinal numbers (e.g., "1st" → "first") | | PLAIN | 377 | 9.6 min | 1.52s | Plain number words | | TELEPHONE | 936 | 61.3 min | 3.93s | Phone numbers | | TIME | 947 | 24.1 min | 1.53s | Time expressions (e.g., "3:00 PM" → "three o'clock p m") | | **TOTAL** | **10,131** | **4.89h** | **1.74s** | | ## Usage ```python from datasets import load_dataset dataset = load_dataset("NNstuff/Numb3rs_NV") ``` ## Metadata Schema | Field | Type | Description | |-------|------|-------------| | `file_name` | string | Relative path to audio file | | `name` | string | Original sample identifier | | `duration` | float | Audio duration in seconds | | `category` | string | Category name (e.g., "MONEY", "DATE") | | `original_text` | string | Written form (TN input) | | `text` | string | Spoken form (ITN input) | | `lang` | string | Language code ("en") | ## NeMo Compatibility For NeMo users, the dataset includes NeMo-format manifests with `audio_filepath` relative paths: - `manifest.jsonl` — Full dataset manifest (resolve paths from repo root) - `manifests/by_category/*.jsonl` — Per-category manifests ## License CC-BY-NC-SA-4.0