README.md

---
license: apache-2.0
task_categories:
  - automatic-speech-recognition
language:
  - en
tags:
  - phoneme-recognition
  - arpabet
  - pronunciation
  - wav2vec2
  - ctc
  - speech
pretty_name: LibriSpeech ARPAbet Phonemes
size_categories:
  - 10K<n<100K
configs:
  - config_name: full
    default: true
    data_files:
      - split: train
        path: data/train/*.parquet
      - split: test
        path: data/test/*.parquet
  - config_name: mini
    data_files:
      - split: train
        path: data/mini/train.parquet
      - split: test
        path: data/mini/test.parquet
dataset_info:
  - config_name: full
    features:
      - name: audio
        dtype: audio
      - name: input_values
        sequence: float32
      - name: labels
        sequence: int32
      - name: text
        dtype: string
      - name: duration
        dtype: float32
    splits:
      - name: train
        num_examples: 15928
      - name: test
        num_examples: 1770
  - config_name: mini
    features:
      - name: audio
        dtype: audio
      - name: input_values
        sequence: float32
      - name: labels
        sequence: int32
      - name: text
        dtype: string
      - name: duration
        dtype: float32
    splits:
      - name: train
        num_examples: 900
      - name: test
        num_examples: 100
---

# LibriSpeech ARPAbet Processed Dataset

Pre-processed dataset for training ARPAbet phoneme recognition models using CTC loss.

## Dataset Description

This dataset is derived from [LibriSpeech](https://huggingface.co/datasets/openslr/librispeech_asr) (train-clean-100 split) with the following preprocessing:

- **Audio**: Resampled to 16kHz, normalized using Wav2Vec2 feature extractor
- **Labels**: Text transcriptions converted to ARPAbet phoneme sequences using CMU Pronouncing Dictionary
- **Filtering**: Samples with out-of-vocabulary words (not in CMU Dict) are excluded
  - Original LibriSpeech train-clean-100: 28,539 samples
  - After filtering: 17,698 samples (62% retained)
  - Skipped: 10,841 samples (38% had at least one OOV word)

### Features

| Feature | Type | Description |
|---------|------|-------------|
| `audio` | `Audio` | Original audio (FLAC, 16kHz) - playable in HF viewer |
| `input_values` | `Sequence[float]` | Normalized audio waveform (float32) - ready for training |
| `labels` | `Sequence[int]` | ARPAbet phoneme token IDs |
| `text` | `string` | Original text transcription |
| `duration` | `float` | Audio duration in seconds |

### ARPAbet Vocabulary (72 tokens)

The vocabulary includes:
- **Special tokens (3)**: `<pad>`, `<unk>`, `|` (word boundary)
- **Consonants (24)**: B, CH, D, DH, F, G, HH, JH, K, L, M, N, NG, P, R, S, SH, T, TH, V, W, Y, Z, ZH
- **Vowels with stress markers (45)**: 15 base vowels x 3 stress levels (0, 1, 2)
  - Example: AA0 (no stress), AA1 (primary), AA2 (secondary)

### Splits

| Split | Samples | Description |
|-------|---------|-------------|
| train | 15,928 | Training data (90%) |
| test  | 1,770  | Evaluation data (10%) |

## Usage

```python
from datasets import load_dataset
from huggingface_hub import hf_hub_download
import json

# Load full dataset (~17k samples, ~15GB)
dataset = load_dataset("davidggphy/librispeech-arpabet-processed", "full")

# Load mini dataset (1000 samples, ~1GB) - great for quick testing!
dataset = load_dataset("davidggphy/librispeech-arpabet-processed", "mini")

# Streaming mode (download samples on-demand)
stream = load_dataset("davidggphy/librispeech-arpabet-processed", "full", split="train", streaming=True)
for sample in stream.take(100):
    print(sample["text"])

# Load vocabulary mapping
vocab_path = hf_hub_download(
    repo_id="davidggphy/librispeech-arpabet-processed",
    filename="vocab.json",
    repo_type="dataset"
)
with open(vocab_path) as f:
    vocab_data = json.load(f)

token_to_id = vocab_data["token_to_id"]
id_to_token = {int(k): v for k, v in vocab_data["id_to_token"].items()}

# Access samples
sample = dataset[0]
print(f"Text: {sample['text']}")
print(f"Duration: {sample['duration']:.2f}s")
print(f"Labels: {[id_to_token[i] for i in sample['labels']]}")
```

### Listening to Audio

The `input_values` column contains normalized audio waveforms at 16kHz. To play or save the audio:

```python
import numpy as np
import soundfile as sf
from IPython.display import Audio

sample = dataset["train"][0]

# Convert to numpy array
audio = np.array(sample["input_values"], dtype=np.float32)

# Play in Jupyter/Colab
Audio(audio, rate=16000)

# Or save to file
sf.write("sample.wav", audio, 16000)

# Check duration
print(f"Duration: {sample['duration']:.2f}s")
print(f"Text: {sample['text']}")
```

### Training with Wav2Vec2

```python
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# Load model and processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base", vocab_size=72)
processor = Wav2Vec2Processor.from_pretrained("davidggphy/wav2vec2-arpabet-phoneme")

# The dataset is ready for CTC training
# input_values: normalized audio
# labels: phoneme token IDs
```

## Intended Use

This dataset is designed for:
- Training phoneme recognition models for English pronunciation assessment
- Fine-tuning Wav2Vec2 for ARPAbet output
- Research in automatic pronunciation evaluation

## Source Data

- **Base Dataset**: [LibriSpeech ASR Corpus](https://huggingface.co/datasets/openslr/librispeech_asr) (train-clean-100 split)
- **Phoneme Dictionary**: [CMU Pronouncing Dictionary](https://github.com/cmusphinx/cmudict)

## Limitations

- Only covers words present in CMU Dictionary (~126k words)
- Based on American English pronunciation
- Does not include phonetic variations or connected speech phenomena

## Citation

If you use this dataset, please cite LibriSpeech:

```bibtex
@inproceedings{panayotov2015librispeech,
  title={Librispeech: an ASR corpus based on public domain audio books},
  author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle={ICASSP},
  year={2015}
}
```

## License

Apache 2.0 (same as LibriSpeech)