Implementation:Speechbrain Speechbrain AMI Diarization Experiment

Knowledge Sources	SpeechBrain
Domains	Speaker_Diarization
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for speaker diarization using deep embeddings and spectral clustering provided by the SpeechBrain library.

Description

This script implements an end-to-end speaker diarization pipeline on the AMI Meeting Corpus. It extracts deep speaker embeddings (e.g., ECAPA-TDNN) from speech segments obtained via Oracle VAD (voice activity detection taken from ground truth), stores them as StatObject_SB objects, and then applies spectral clustering (or optionally other backends) to assign speaker labels. The pipeline iterates over individual recordings to reduce GPU memory demands. The final output is an RTTM-formatted speaker boundary file, and the system is evaluated using the Diarization Error Rate (DER) metric.

The recipe is based on the paper: N. Dawalatabad, M. Ravanelli, F. Grondin, J. Thienpondt, B. Desplanques, H. Na, "ECAPA-TDNN Embeddings for Speaker Diarization," arXiv:2104.01466, 2021.

Usage

Run as a standalone recipe script with a YAML hyperparameter file specifying the embedding model, clustering parameters, and dataset paths. Typical usage is for evaluating diarization performance on the AMI corpus with Oracle VAD conditions.

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/AMI/Diarization/experiment.py

Signature

def compute_embeddings(wavs, lens):
    """Definition of the steps for computation of embeddings from the waveforms."""
    ...

def embedding_computation_loop(split, set_loader, stat_file):
    """Extracts embeddings for a given dataset loader."""
    ...

def prepare_subset_json(full_meta_data, rec_id, out_meta_file):
    """Prepares metadata for a given recording ID."""
    ...

def diarize_dataset(full_meta, split_type, n_lambdas, pval, n_neighbors=10):
    """Diarizes all the recordings in a given dataset using spectral clustering."""
    ...

Import

python experiment.py hparams/ecapa_tdnn.yaml

I/O Contract

Inputs

Name	Type	Required	Description
hparams_file	str	Yes	Path to YAML hyperparameter file (e.g., ecapa_tdnn.yaml)
wavs	torch.Tensor	Yes	Waveform tensor batch for embedding extraction
lens	torch.Tensor	Yes	Relative lengths of each waveform in the batch
full_meta	dict	Yes	Full JSON metadata containing all recordings
split_type	str	Yes	Dataset split identifier (e.g., "dev", "eval")
n_lambdas	int	Yes	Number of eigenvalues for spectral clustering tuning
pval	float	Yes	p-value threshold for spectral clustering

Outputs

Name	Type	Description
RTTM file	file	Speaker boundary file in RTTM format per recording
DER scores	float	Diarization Error Rate computed against ground truth
stat_obj	StatObject_SB	Stored embeddings object with modelset, segset, and stat1

Usage Examples

# Run the diarization experiment with ECAPA-TDNN embeddings
python experiment.py hparams/ecapa_tdnn.yaml

# With overrides
python experiment.py hparams/ecapa_tdnn.yaml --data_folder /path/to/AMI --output_folder results/

Related Pages

• Principle:Speechbrain_Speechbrain_Speaker_Diarization_Pipeline