Edit model card

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)

GitHub GitHub

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.

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
Downloads last month
170
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train Jzuluaga/wav2vec2-xls-r-300m-en-atc-atcosim

Evaluation results