Upload README.md with huggingface_hub

README.md

@@ -20,6 +20,12 @@ size_categories:
 
 Pre-processed dataset for training ARPAbet phoneme recognition models using CTC loss.
 
+## Inspiration
+
+This project was inspired by Simon Edwards' blog post on pronunciation training using CTC:
+- [Ear Pronunciation via CTC](https://simedw.com/2026/01/31/ear-pronunication-via-ctc/) - Blog post
+- [Hacker News Discussion](https://news.ycombinator.com/item?id=46832074)
+
 ## Dataset Description
 
 This dataset is derived from [LibriSpeech](https://huggingface.co/datasets/librispeech_asr) (train-clean-100 split) with the following preprocessing:
@@ -54,14 +60,28 @@ The vocabulary includes:
 
 ```python
 from datasets import load_dataset
+from huggingface_hub import hf_hub_download
+import json
 
 # Load the dataset
 dataset = load_dataset("davidggphy/librispeech-arpabet-processed")
 
+# 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["train"][0]
-print(f"Audio shape: {len(sample['input_values'])}")
-print(f"Labels: {sample['labels']}")
+print(f"Audio length: {len(sample['input_values'])} samples")
+print(f"Labels: {[id_to_token[i] for i in sample['labels']]}")
 ```
 
 ### Training with Wav2Vec2