miguelozaalon/speaker-segmentation-atc

Speaker Segmentation for Air Traffic Control Communications

This model is a fine-tuned version of pyannote/segmentation-3.0, adapted to speaker segmentation in air traffic control (ATC) communications.

Model description

The model predicts speaker activity and segmentation in ATC audio and can be used as a component of speaker diarization and transcription pipelines.

It was fine-tuned on ATC-specific audio derived from the ATCO2-test-set-1h corpus.

Intended uses & limitations

Intended uses

• Speaker segmentation in air traffic control recordings
• Use as a component in ATC speaker diarization pipelines
• Research on transcription and analysis of ATC communications
• Analysis of speaker turn-taking patterns

Training and evaluation data

The model was fine-tuned using a speaker-segmentation dataset derived from the ATCO2-test-set-1h corpus.

The dataset contains ATC communications between air traffic controllers and pilots under varied acoustic conditions and was split into training and validation subsets.

Training procedure

Fine-tuning started from the pretrained pyannote/segmentation-3.0 model and adapted it to the acoustic and conversational characteristics of air traffic control communications.

Training hyperparameters

• Learning rate: 0.001
• Train batch size: 32
• Evaluation batch size: 32
• Seed: 42
• Optimizer: Adam, betas=(0.9, 0.999), epsilon=1e-08
• Learning-rate scheduler: cosine
• Epochs: 5

Evaluation results

Evaluation setting	DER ↓	JER ↓
ATCO2	15.816%	24.198%
ATCO2 + noise reduction	14.764%	19.815%

Framework versions

• Transformers 4.45.1
• PyTorch 2.4.1+cu124
• Datasets 3.0.1
• Tokenizers 0.20.0

License and access

The base model, pyannote/segmentation-3.0, is licensed under the MIT License.

This model was fine-tuned using the ATCO2-test-set-1h corpus, which is subject to the ATCO2 Data End-User Agreement and restricted to research use. Because of the distribution restrictions associated with the training data, the fine-tuned model weights are not made openly available.

ATCO2 attribution

Use of the ATCO2 data is subject to the ATCO2 Data End-User Agreement. Users should refer to https://www.atco2.org/ and cite the publication required by the ATCO2 agreement.