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

Training Loss	Epoch	Step	Validation Loss	Wer
1.4757	6.41	500	0.0614	0.0347
0.0624	12.82	1000	0.0525	0.0277
0.0388	19.23	1500	0.0693	0.0241
0.03	25.64	2000	0.0666	0.0244
0.0235	32.05	2500	0.0604	0.0260
0.0226	38.46	3000	0.0625	0.0230
0.0163	44.87	3500	0.0603	0.0195
0.0157	51.28	4000	0.0628	0.0209
0.0152	57.69	4500	0.0692	0.0238
0.0122	64.1	5000	0.0607	0.0210
0.011	70.51	5500	0.0608	0.0213
0.0114	76.92	6000	0.0681	0.0211
0.0106	83.33	6500	0.0613	0.0210
0.0081	89.74	7000	0.0654	0.0196
0.0078	96.15	7500	0.0612	0.0191
0.0082	102.56	8000	0.0758	0.0237
0.0078	108.97	8500	0.0664	0.0206
0.0075	115.38	9000	0.0658	0.0197
0.0052	121.79	9500	0.0669	0.0218
0.0054	128.21	10000	0.0695	0.0211
0.0053	134.62	10500	0.0726	0.0227
0.0046	141.03	11000	0.0702	0.0212
0.0043	147.44	11500	0.0846	0.0200
0.0041	153.85	12000	0.0764	0.0200
0.0032	160.26	12500	0.0785	0.0201
0.0028	166.67	13000	0.0839	0.0197
0.0035	173.08	13500	0.0785	0.0210
0.0027	179.49	14000	0.0730	0.0188
0.002	185.9	14500	0.0794	0.0193
0.002	192.31	15000	0.0859	0.0211
0.0019	198.72	15500	0.0727	0.0183
0.0017	205.13	16000	0.0784	0.0187
0.0016	211.54	16500	0.0801	0.0196
0.0014	217.95	17000	0.0821	0.0185
0.0011	224.36	17500	0.0822	0.0176
0.001	230.77	18000	0.0856	0.0171
0.001	237.18	18500	0.0792	0.0176
0.001	243.59	19000	0.0826	0.0173
0.0006	250.0	19500	0.0854	0.0170
0.0007	256.41	20000	0.0850	0.0167

Framework versions

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