Implementation:Speechbrain Speechbrain Prepare AMI Diarization

Knowledge Sources	SpeechBrain
Domains	Speaker Diarization, Data Preparation
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for preparing AMI corpus data for speaker diarization provided by the SpeechBrain library.

Description

This script prepares metadata files (JSON) and reference RTTM files from the AMI Meeting Corpus manual annotations for speaker diarization tasks. It processes XML-based segment annotations into RTTM format (Oracle VAD), splits the data according to standard AMI corpus splits (scenario_only, full_corpus, or full_corpus_asr), and generates sub-segments with configurable duration and overlap for diarization model training and evaluation.

Usage

Use this when preparing the AMI Meeting Corpus dataset for speaker diarization training with SpeechBrain recipes.

Code Reference

Source Location

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

Signature

def prepare_ami(
    data_folder,
    manual_annot_folder,
    save_folder,
    ref_rttm_dir,
    meta_data_dir,
    split_type="full_corpus_asr",
    skip_TNO=True,
    mic_type="Lapel",
    vad_type="oracle",
    max_subseg_dur=3.0,
    overlap=1.5,
):

Import

from ami_prepare import prepare_ami

I/O Contract

Inputs

Name	Type	Required	Description
data_folder	str	Yes	Path to the folder where the original AMI corpus is stored
manual_annot_folder	str	Yes	Directory where the manual annotations are stored
save_folder	str	Yes	The save directory in results
ref_rttm_dir	str	Yes	Directory to store reference RTTM files
meta_data_dir	str	Yes	Directory to store the metadata (JSON) files
split_type	str	No	Standard dataset split: "scenario_only", "full_corpus", or "full_corpus_asr" (default: "full_corpus_asr")
skip_TNO	bool	No	Skips TNO meeting recordings if True (default: True)
mic_type	str	No	Type of microphone to be used (default: "Lapel")
vad_type	str	No	Type of VAD (default: "oracle")
max_subseg_dur	float	No	Maximum sub-segment duration in seconds (default: 3.0)
overlap	float	No	Overlap between sub-segments in seconds (default: 1.5)

Outputs

Name	Type	Description
RTTM files	RTTM Files	Reference RTTM files for train/dev/test splits
JSON metadata	JSON Files	Prepared metadata manifest files for train/dev/test splits

Usage Examples

from ami_prepare import prepare_ami

prepare_ami(
    data_folder="/path/to/amicorpus",
    manual_annot_folder="/path/to/ami_manual_annotations",
    save_folder="/path/to/output",
    ref_rttm_dir="/path/to/output/ref_rttms",
    meta_data_dir="/path/to/output/metadata",
    split_type="full_corpus_asr",
    mic_type="Lapel",
    max_subseg_dur=3.0,
    overlap=1.5,
)

Related Pages

• Principle:Speechbrain_Speechbrain_Dataset_Specific_Data_Preparation