metadata

license: mit
base_model: pyannote/segmentation-3.0
tags:
  - speaker-diarization
  - speaker-segmentation
  - generated_from_trainer
datasets:
  - diarizers-community/callhome
model-index:
  - name: speaker-segmentation-fine-tuned-callhome-jpn
    results: []

speaker-segmentation-fine-tuned-callhome-jpn

This model is a fine-tuned version of pyannote/segmentation-3.0 on the diarizers-community/callhome jpn dataset. It achieves the following results on the evaluation set:

Loss: 0.7653
Der: 0.2311
False Alarm: 0.0477
Missed Detection: 0.1352
Confusion: 0.0482

Model description

This segmentation model has been trained on Japanese data (Callhome) using diarizers. It can be loaded with two lines of code:

from diarizers import SegmentationModel

segmentation_model = SegmentationModel().from_pretrained('diarizers-community/speaker-segmentation-fine-tuned-callhome-jpn')

To use it within a pyannote speaker diarization pipeline, load the pyannote/speaker-diarization-3.1 pipeline, and convert the model to a pyannote compatible format:


from pyannote.audio import Pipeline
import torch

device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")

# load the pre-trained pyannote pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")
pipeline.to(device)

# replace the segmentation model with your fine-tuned one
segmentation_model = segmentation_model.to_pyannote_model()
pipeline._segmentation.model = segmentation_model.to(device)

You can now use the pipeline on audio examples:

from datasets import load_dataset
# load dataset example
dataset = load_dataset("diarizers-community/callhome", "jpn", split="data")
sample = dataset[0]["audio"]

# pre-process inputs
sample["waveform"] = torch.from_numpy(sample.pop("array")[None, :]).to(device, dtype=model.dtype)
sample["sample_rate"] = sample.pop("sampling_rate")

# perform inference
diarization = pipeline(sample)

# dump the diarization output to disk using RTTM format
with open("audio.rttm", "w") as rttm:
    diarization.write_rttm(rttm)

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.001
train_batch_size: 32
eval_batch_size: 32
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
num_epochs: 5.0

Training results

Training Loss	Epoch	Step	Validation Loss	Der	False Alarm	Missed Detection	Confusion
0.5917	1.0	328	0.7859	0.2409	0.0507	0.1369	0.0533
0.5616	2.0	656	0.7738	0.2350	0.0530	0.1350	0.0471
0.5364	3.0	984	0.7737	0.2358	0.0484	0.1368	0.0506
0.5121	4.0	1312	0.7626	0.2317	0.0483	0.1358	0.0475
0.5166	5.0	1640	0.7653	0.2311	0.0477	0.1352	0.0482

Framework versions

Transformers 4.40.0
Pytorch 2.2.2+cu121
Datasets 2.18.0
Tokenizers 0.19.1