|
# AMI Corpus |
|
|
|
https://groups.inf.ed.ac.uk/ami/corpus/ |
|
|
|
To be filled! |
|
|
|
**Note**: This dataset corresponds to the data-processing of [KALDI's AMI S5 recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/ami/s5). |
|
This means text is normalized and the audio data is chunked according to the scripts above! |
|
To make the user experience as simply as possible, we provide the already chunked data to the user here so that the following can be done: |
|
|
|
```python |
|
from datasets import load_dataset |
|
ds = load_dataset("edinburghcstr/ami", "ihm") |
|
|
|
print(ds) |
|
``` |
|
gives: |
|
``` |
|
DatasetDict({ |
|
train: Dataset({ |
|
features: ['segment_id', 'audio_id', 'text', 'audio', 'begin_time', 'end_time', 'microphone_id', 'speaker_id'], |
|
num_rows: 108502 |
|
}) |
|
validation: Dataset({ |
|
features: ['segment_id', 'audio_id', 'text', 'audio', 'begin_time', 'end_time', 'microphone_id', 'speaker_id'], |
|
num_rows: 13098 |
|
}) |
|
test: Dataset({ |
|
features: ['segment_id', 'audio_id', 'text', 'audio', 'begin_time', 'end_time', 'microphone_id', 'speaker_id'], |
|
num_rows: 12643 |
|
}) |
|
}) |
|
``` |
|
|
|
```py |
|
ds["train"][0] |
|
``` |
|
|
|
automatically loads the audio into memory: |
|
|
|
``` |
|
{'segment_id': 'EN2001a', |
|
'audio_id': 'AMI_EN2001a_H00_MEE068_0000557_0000594', |
|
'text': 'OKAY', |
|
'audio': {'path': '/home/patrick_huggingface_co/.cache/huggingface/datasets/downloads/extracted/2d75d5b3e8a91f44692e2973f08b4cac53698f92c2567bd43b41d19c313a5280/EN2001a/train_ami_en2001a_h00_mee068_0000557_0000594.wav', |
|
'array': array([0. , 0. , 0. , ..., 0.00033569, 0.00030518, |
|
0.00030518], dtype=float32), |
|
'sampling_rate': 16000}, |
|
'begin_time': 5.570000171661377, |
|
'end_time': 5.940000057220459, |
|
'microphone_id': 'H00', |
|
'speaker_id': 'MEE068'} |
|
``` |
|
|
|
|
|
The dataset was tested for correctness by fine-tuning a Wav2Vec2-Large model on it, more explicitly [the `wav2vec2-large-lv60` checkpoint](https://huggingface.co/facebook/wav2vec2-large-lv60). |
|
|
|
As can be seen in this experiments, training the model for less than 2 epochs gives |
|
|
|
*Result (WER)*: |
|
|
|
| "dev" | "eval" | |
|
|---|---| |
|
| 25.27 | 25.21 | |
|
|
|
as can be seen [here](https://huggingface.co/patrickvonplaten/ami-wav2vec2-large-lv60). |
|
|
|
The results are in-line with results of published papers: |
|
|
|
- [*Hybrid acoustic models for distant and multichannel large vocabulary speech recognition*](https://www.researchgate.net/publication/258075865_Hybrid_acoustic_models_for_distant_and_multichannel_large_vocabulary_speech_recognition) |
|
- [Multi-Span Acoustic Modelling using Raw Waveform Signals](https://arxiv.org/abs/1906.11047) |
|
|
|
You can run [run.sh](https://huggingface.co/patrickvonplaten/ami-wav2vec2-large-lv60/blob/main/run.sh) to reproduce the result. |
