wav2vec2-large-960h-lv60-self-en-atc-atcosim

This model is a fine-tuned version of facebook/wav2vec2-large-960h-lv60-self on the ATCOSIM corpus.

It achieves the following results on the evaluation set:

• Loss: 0.0850
• Wer: 0.0167 (1.67% WER)

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

(you need to change the MODEL_ID param to MODEL_ID=Jzuluaga/wav2vec2-large-960h-lv60-self-en-atc-atcosim)

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-large-960h-lv60-self-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

Framework versions

• Transformers 4.24.0
• Pytorch 1.13.0+cu117
• Datasets 2.6.1
• Tokenizers 0.13.2