metadata

language:
  - en
license: apache-2.0
base_model: openai/whisper-small
tags:
  - speaker-diarization
  - speaker-segmentation
  - generated_from_trainer
model-index:
  - name: speaker-segmentation-eng
    results: []

speaker-segmentation-eng

This model is a fine-tuned version of openai/whisper-small on the diarizers-community/callhome dataset. It achieves the following results on the evaluation set:

Loss: 0.4666
Der: 0.1827
False Alarm: 0.0590
Missed Detection: 0.0715
Confusion: 0.0522

Model description

This segmentation model has been trained on English 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 diarizers import SegmentationModel
from pyannote.audio import Pipeline
from datasets import load_dataset
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)

model = SegmentationModel().from_pretrained("nehulagrawal/speaker-segmentation-eng")
model = model.to_pyannote_model()
pipeline._segmentation.model = 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", "eng", 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)

You can now use the pipeline on single audio examples:


from diarizers import SegmentationModel
from pyannote.audio import Pipeline
from datasets import load_dataset
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)

model = SegmentationModel().from_pretrained("nehulagrawal/speaker-segmentation-eng")
model = model.to_pyannote_model()
pipeline._segmentation.model = model.to(device)

diarization = pipeline("audio.wav")
with open("audio.rttm", "w") as rttm:
    diarization.write_rttm(rttm)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.001
train_batch_size: 64
eval_batch_size: 64
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss	Der	False Alarm	Missed Detection	Confusion
0.4224	1.0	181	0.4837	0.1939	0.0599	0.0764	0.0576
0.409	2.0	362	0.4692	0.1884	0.0618	0.0724	0.0543
0.3919	3.0	543	0.4700	0.1875	0.0638	0.0698	0.0540
0.3693	4.0	724	0.4718	0.1848	0.0602	0.0714	0.0533
0.358	5.0	905	0.4606	0.1810	0.0544	0.0754	0.0512
0.355	6.0	1086	0.4631	0.1826	0.0638	0.0677	0.0512
0.3563	7.0	1267	0.4646	0.1809	0.0587	0.0716	0.0505
0.347	8.0	1448	0.4682	0.1820	0.0581	0.0720	0.0519
0.3463	9.0	1629	0.4684	0.1827	0.0586	0.0718	0.0523
0.3299	10.0	1810	0.4666	0.1827	0.0590	0.0715	0.0522

Framework versions

Transformers 4.40.1
Pytorch 2.2.1+cu121
Datasets 2.19.1
Tokenizers 0.19.1