espnet/diar_ami_eend
End-to-end Neural Diarization (EEND) trained on AMI-headset dataset. 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.
- Follow the frame concatenation and subsampling strategy described in paper [[2]]. Each frame is concatenated with the preceding and following 7 frames, followed by subsampling with a factor of 10. As a result, a 345-dimensional acoustic feature (23 × 15) is extracted for each 100 ms.
- Training and testing are performed exclusively on data with 4 speakers.
- Use 4 layer stacked Transformer encoder, each outputs 256-dimensional frame-wise embeddings.
- The training process spans 500 epochs.
- Detailed configurations are defined in
exp/diar/train_diar_diar_raw/config.yaml.
RESULTS
Environments
- date:
Thu Dec 19 22:03:53 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
diartraindiar_raw
DER
diarizedtest |thresholdmediancollar|DER| |---|---| |resultth0.3med11collar0.0|71.73| |resultth0.3med1collar0.0|74.62| |resultth0.4med11collar0.0|70.10| |resultth0.4med1collar0.0|71.98| |resultth0.5med11collar0.0|70.57| |resultth0.5med1collar0.0|72.44| |resultth0.6med11collar0.0|72.64| |resultth0.6med1collar0.0|74.63| |resultth0.7med11collar0.0|76.52| |resultth0.7med1_collar0.0|78.41|
diartraindiar_raw
DER
diarizeddev |thresholdmediancollar|DER| |---|---| |resultth0.3med11collar0.0|75.88| |resultth0.3med1collar0.0|78.21| |resultth0.4med11collar0.0|71.45| |resultth0.4med1collar0.0|73.32| |resultth0.5med11collar0.0|70.53| |resultth0.5med1collar0.0|72.34| |resultth0.6med11collar0.0|72.03| |resultth0.6med1collar0.0|73.96| |resultth0.7med11collar0.0|76.66| |resultth0.7med1_collar0.0|78.33|