wav2vec2-xls-r-300m-en-atc-atcosim

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the ATCOSIM corpus.

(A better ASR model for ATC data is available here: https://huggingface.co./Jzuluaga/wav2vec2-xls-r-300m-en-atc-uwb-atcc-and-atcosim)

It achieves the following results on the evaluation set:

Loss: 0.0988
Wer: 0.0736

Paper: How Does Pre-trained Wav2Vec 2.0 Perform on Domain Shifted ASR? An Extensive Benchmark on Air Traffic Control Communications.

Authors: Juan Zuluaga-Gomez, Amrutha Prasad, Iuliia Nigmatulina, Saeed Sarfjoo, Petr Motlicek, Matthias Kleinert, Hartmut Helmke, Oliver Ohneiser, Qingran Zhan

Abstract: Recent work on self-supervised pre-training focus on leveraging large-scale unlabeled speech data to build robust end-to-end (E2E)acoustic models (AM) that can be later fine-tuned on downstream tasks e.g., automatic speech recognition (ASR). Yet, few works investigated the impact on performance when the data properties substantially differ between the pre-training and fine-tuning phases, termed domain shift. We target this scenario by analyzing the robustness of Wav2Vec 2.0 and XLS-R models on downstream ASR for a completely unseen domain, air traffic control (ATC) communications. We benchmark these two models on several open-source and challenging ATC databases with signal-to-noise ratio between 5 and 20 dB. Relative word error rate (WER) reductions between 20% to 40% are obtained in comparison to hybrid-based ASR baselines by only fine-tuning E2E acoustic models with a smaller fraction of labeled data. We analyze WERs on the low-resource scenario and gender bias carried by one ATC dataset.

Code — GitHub repository: https://github.com/idiap/w2v2-air-traffic

Usage

You can use our Google Colab notebook to run and evaluate our model: https://github.com/idiap/w2v2-air-traffic/blob/master/src/eval_xlsr_atc_model.ipynb

Intended uses & limitations

This model was fine-tuned on air traffic control data. We don't expect that it keeps the same performance on some others datasets, e.g., LibriSpeech or CommonVoice.

Training and evaluation data

See Table 1 (page 3) in our paper: How Does Pre-trained Wav2Vec 2.0 Perform on Domain Shifted ASR? An Extensive Benchmark on Air Traffic Control Communications. We described there the partitions of how to use our model.

We use the ATCOSIM dataset for fine-tuning this model. You can download the raw data here: https://www.spsc.tugraz.at/databases-and-tools/atcosim-air-traffic-control-simulation-speech-corpus.html
However, do not worry, we have prepared the database in Datasets format. Here, ATCOSIM CORPUS on HuggingFace. You can scroll and check the train/test partitions, and even listen to some audios.
If you want to prepare a database in HuggingFace format, you can follow the data loader script in: data_loader_atc.py.

Writing your own inference script

If you use language model, you need to install the KenLM bindings with:

conda activate your_environment
pip install https://github.com/kpu/kenlm/archive/master.zip

The snippet of code:

from datasets import load_dataset, load_metric, Audio
import torch
from transformers import AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM
import torchaudio.functional as F

USE_LM = False
DATASET_ID = "Jzuluaga/atcosim_corpus"
MODEL_ID = "Jzuluaga/wav2vec2-xls-r-300m-en-atc-atcosim"

# 1. Load the dataset
# we only load the 'test' partition, however, if you want to load the 'train' partition, you can change it accordingly
atcosim_corpus_test = load_dataset(DATASET_ID, "test", split="test")

# 2. Load the model
model = AutoModelForCTC.from_pretrained(MODEL_ID)

# 3. Load the processors, we offer support with LM, which should yield better resutls
if USE_LM:
    processor = Wav2Vec2ProcessorWithLM.from_pretrained(MODEL_ID)
else:
    processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)

# 4. Format the test sample
sample = next(iter(atcosim_corpus_test))
file_sampling_rate = sample['audio']['sampling_rate']

# resample if neccessary
if file_sampling_rate != 16000:
    resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), file_sampling_rate, 16000).numpy()
else:
    resampled_audio = torch.tensor(sample["audio"]["array"]).numpy()

input_values = processor(resampled_audio, return_tensors="pt").input_values

# 5. Run the forward pass in the model
with torch.no_grad():
    logits = model(input_values).logits
    
# get the transcription with processor
if USE_LM:
    transcription = processor.batch_decode(logits.numpy()).text
else:
    pred_ids = torch.argmax(logits, dim=-1)
    transcription = processor.batch_decode(pred_ids)

# print the output
print(transcription)

Cite us

If you use this code for your research, please cite our paper with:

@article{zuluaga2022how,
    title={How Does Pre-trained Wav2Vec2. 0 Perform on Domain Shifted ASR? An Extensive Benchmark on Air Traffic Control Communications},
    author={Zuluaga-Gomez, Juan and Prasad, Amrutha and Nigmatulina, Iuliia and Sarfjoo, Saeed and others},
    journal={IEEE Spoken Language Technology Workshop (SLT), Doha, Qatar},
    year={2022}
  }

and,

@article{zuluaga2022bertraffic,
  title={BERTraffic: BERT-based Joint Speaker Role and Speaker Change Detection for Air Traffic Control Communications},
  author={Zuluaga-Gomez, Juan and Sarfjoo, Seyyed Saeed and Prasad, Amrutha and others},
  journal={IEEE Spoken Language Technology Workshop (SLT), Doha, Qatar},
  year={2022}
  }

and,

@article{zuluaga2022atco2,
  title={ATCO2 corpus: A Large-Scale Dataset for Research on Automatic Speech Recognition and Natural Language Understanding of Air Traffic Control Communications},
  author={Zuluaga-Gomez, Juan and Vesel{\`y}, Karel and Sz{\"o}ke, Igor and Motlicek, Petr and others},
  journal={arXiv preprint arXiv:2211.04054},
  year={2022}
}

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0005
train_batch_size: 24
eval_batch_size: 24
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 96
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
training_steps: 20000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.9105	6.41	500	0.1622	0.1531
0.1119	12.82	1000	0.0971	0.0936
0.0614	19.23	1500	0.1002	0.0983
0.044	25.64	2000	0.1011	0.0929
0.0366	32.05	2500	0.0932	0.0828
0.0315	38.46	3000	0.0926	0.0880
0.0297	44.87	3500	0.0972	0.0882
0.0216	51.28	4000	0.0911	0.0774
0.0211	57.69	4500	0.0982	0.0891
0.0187	64.1	5000	0.1009	0.0863
0.02	70.51	5500	0.0953	0.0852
0.0163	76.92	6000	0.1028	0.0804
0.0128	83.33	6500	0.0930	0.0856
0.0127	89.74	7000	0.0892	0.0676
0.0116	96.15	7500	0.0857	0.0753
0.0139	102.56	8000	0.1078	0.0481
0.0107	108.97	8500	0.0955	0.0683
0.0096	115.38	9000	0.0846	0.0697
0.0089	121.79	9500	0.0854	0.0675
0.0084	128.21	10000	0.0875	0.0779
0.0074	134.62	10500	0.0840	0.0770
0.0061	141.03	11000	0.0903	0.0754
0.0076	147.44	11500	0.0872	0.0769
0.0069	153.85	12000	0.0891	0.0772
0.0061	160.26	12500	0.0971	0.0774
0.0049	166.67	13000	0.0984	0.0726
0.0045	173.08	13500	0.0952	0.0765
0.0039	179.49	14000	0.1015	0.0762
0.0031	185.9	14500	0.0937	0.0712
0.0032	192.31	15000	0.0982	0.0635
0.0028	198.72	15500	0.0981	0.0743
0.0024	205.13	16000	0.1019	0.0712
0.0024	211.54	16500	0.0957	0.0732
0.002	217.95	17000	0.0941	0.0732
0.0015	224.36	17500	0.1009	0.0717
0.0017	230.77	18000	0.0955	0.0730
0.0013	237.18	18500	0.0989	0.0732
0.0013	243.59	19000	0.0967	0.0738
0.0011	250.0	19500	0.0980	0.0734
0.0008	256.41	20000	0.0988	0.0736

Framework versions

Transformers 4.24.0
Pytorch 1.13.0+cu117
Datasets 2.6.1
Tokenizers 0.13.2

Jzuluaga
/

wav2vec2-xls-r-300m-en-atc-atcosim