diar_ami_eend_eda

End-to-end Neural Diarization with Encoder-Decoder Based Attractors trained on AMI-headset. This example could be found at egs2/ami/diar1.

Configurations:

• Use ESPNet's default frontend to extract features. The sampling rate is 8000 Hz, with a frame length of 25 ms and a frame shift of 10 ms. The frontend extracts 23 log-scaled Mel-filterbanks.
• Use 4 layer stacked Transformer encoder, each outputs 256-dimensional frame-wise embeddings.
• Use the ESPNet' standard rnn attractor (LSTM) with hidden size of 256.
• Initial training uses data with 4 speakers for 500 epochs, following spk4/diar_train_diar_eda_raw_spk4/config.yaml.
• Adaptation involves fine-tuning the model using data with 3 and 5 speakers respectively for 20 epochs respectively, using spk3/diar_train_diar_eda_raw_spk3/config.yaml and spk5/diar_train_diar_eda_raw_spk5/config.yaml respectively.

RESULTS

The following results were obtained using the checkpoint spk5/diar_train_diar_eda_raw_spk5/20epoch.pth, tested on the test and development sets with the 4-speakers.

Environments

• date: Thu Dec 19 22:43:37 EST 2024
• python version: 3.11.10 (main, Oct 3 2024, 07:29:13) [GCC 11.2.0]
• espnet version: espnet 202409
• pytorch version: pytorch 2.4.0
• Git hash: c12b3d59ca4fd8847edf274e56a1716474d2a30e
  • Commit date: Thu Dec 19 21:58:26 2024 -0500

spk4

DER

diarized_test

spk4

DER

diarized_dev