|
--- |
|
license: apache-2.0 |
|
language_creators: |
|
- expert-generated |
|
task_categories: |
|
- text-generation |
|
tags: |
|
- generative error correction |
|
- large language model |
|
- LLaMA |
|
pretty_name: Robust HyPoradise |
|
size_categories: |
|
- 100K<n<1M |
|
language: |
|
- en |
|
--- |
|
|
|
# HypothesesParadise |
|
This repo releases the Robust HyPoradise dataset in paper "Large Language Models are Efficient Learners of Noise-Robust Speech Recognition." |
|
|
|
**GitHub:** https://github.com/YUCHEN005/RobustGER |
|
|
|
**Model:** https://huggingface.co/PeacefulData/RobustGER |
|
|
|
**Data:** This repo |
|
|
|
|
|
**UPDATE (Apr-18-2024):** We have released the training data, which follows the same format as test data. |
|
Considering the file size, the uploaded training data does not contain the speech features (vast size). |
|
Alternatively, we have provided a script named ***add_speech_feats_to_train_data.py*** to generate them from raw speech (.wav). |
|
You need to specify the raw speech path from utterance id in the script. |
|
Here are the available speech data: [CHiME-4](https://entuedu-my.sharepoint.com/:f:/g/personal/yuchen005_e_ntu_edu_sg/EuLgMQbjrIJHk7dKPkjcDMIB4SYgXKKP8VBxyiZk3qgdgA), |
|
[VB-DEMAND](https://datashare.ed.ac.uk/handle/10283/2791), [LS-FreeSound](https://github.com/archiki/Robust-E2E-ASR), [NOIZEUS](https://ecs.utdallas.edu/loizou/speech/noizeus/). |
|
|
|
|
|
**IMPORTANT:** The vast speech feature size mentioned above is because Whisper requires a fix input length of 30s that is too long. Please do the follwing step to remove it before running ***add_speech_feats_to_train_data.py***: |
|
- Modified the [whisper model code](https://github.com/openai/whisper/blob/main/whisper/model.py#L167) `x = (x + self.positional_embedding).to(x.dtype)` to be `x = (x + self.positional_embedding[:x.shape[1], :]).to(x.dtype)` |
|
|
|
|
|
**UPDATE (Apr-29-2024):** To support customization, We release the script ***generate_robust_hp.py*** for users to generate train/test data from their own ASR datasets. |
|
We also release two necessary packages for generation: "my_jiwer" and "decoding.py". |
|
To summary, you will need to do the following three steps before running ***generate_robust_hp.py***: |
|
- Modified the [whisper model code](https://github.com/openai/whisper/blob/main/whisper/model.py#L167) `x = (x + self.positional_embedding).to(x.dtype)` to be `x = (x + self.positional_embedding[:x.shape[1], :]).to(x.dtype)` |
|
- Specify the absolute path of "my_jiwer" directory in ***generate_robust_hp.py*** (`sys.path.append()`) |
|
- Put our whisper decoding script "decoding.py" under your locally installed whisper directory "\<your-path\>/whisper/whisper" |
|
|
|
|
|
If you consider this work would be related or useful for your research, please kindly consider to cite the work in ICLR 2024. Thank you. |
|
|
|
```bib |
|
@inproceedings{hu2024large, |
|
title={Large Language Models are Efficient Learners of Noise-Robust Speech Recognition}, |
|
author={Hu, Yuchen and Chen, Chen and Yang, Chao-Han Huck and Li, Ruizhe and Zhang, Chao and Chen, Pin-Yu and Chng, Eng Siong}, |
|
booktitle={International Conference on Learning Representations}, |
|
year={2024} |
|
} |
|
``` |