README.md

---
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: label
    dtype:
      class_label:
        names:
          '0': I+have+one+now
          '1': I+only+have+one
  splits:
  - name: train
    num_bytes: 10168367.5
    num_examples: 535
  - name: test
    num_bytes: 1499291.5
    num_examples: 95
  - name: validation
    num_bytes: 1720511.5
    num_examples: 97
  download_size: 13330229
  dataset_size: 13388170.5
---
# Dataset Card for "have_one"
The dataset consists of utterances of *have one* that are cut either from an utterance of *I have one now*, or from an utterance
of *I only have one*.  The first tends to have prominence on *have*, while the second tends to have prominence on *one*. See
`github.com/MatsRooth/fiyou` on the methodology for finding the utterances on Youtube, and aligning and cutting them using
Kaldi.

To put such a dataset on huggingface hub, start with this directory structure, where the bottom directories contain wav files.  
```
have_one
└── data
    ├── I+have+one+now
    └── I+only+have+one
```
Run `have_one_hub.py` to create the dataset, using the generic Huggingface methodology for audio datasets. 

The dataset is used in the wav2vec2 binary classification model `MatsRooth/wav2vec2-base_have_one`.

Often cutting with a Kaldi phone alignment gives a snippet that includes part of preceding vowel, or has formant structure
in the start of /h/ that gives information about the preceding vowel. These vowels are different for
the two classes, and so classification can be based on this, as well as the intended prosodic difference.
This needs to be corrected.