A Semi-Automatic Approach to Create Large Gender- and Age-Balanced Speaker Corpora: Usefulness of Speaker Diarization & Identification

TL;DR

This work addresses the need for large, diachronically balanced speaker corpora with known gender and age attributes. It introduces a semi-automatic pipeline that combines Clean Speech Detection, VAD/OVL, NSE-based noise handling, VBx diarization, manual speaker identification, and cross-document x-vector matching to assemble a target set of speakers from INA archives. The approach yields 874 speakers across $32$ categories (2 genders, 4 age groups, 4 periods) with substantial manual-time savings and a 3+ minute per speaker requirement, while providing both objective diarization metrics (DER) and subjective quality assessments. The resulting methodology and dataset aim to enable sociolinguistic analyses of gender representation in broadcast media and offer reusable tools for scalable corpus creation with minimal supervision.

Abstract

This paper presents a semi-automatic approach to create a diachronic corpus of voices balanced for speaker's age, gender, and recording period, according to 32 categories (2 genders, 4 age ranges and 4 recording periods). Corpora were selected at French National Institute of Audiovisual (INA) to obtain at least 30 speakers per category (a total of 960 speakers; only 874 have be found yet). For each speaker, speech excerpts were extracted from audiovisual documents using an automatic pipeline consisting of speech detection, background music and overlapped speech removal and speaker diarization, used to present clean speaker segments to human annotators identifying target speakers. This pipeline proved highly effective, cutting down manual processing by a factor of ten. Evaluation of the quality of the automatic processing and of the final output is provided. It shows the automatic processing compare to up-to-date process, and that the output provides high quality speech for most of the selected excerpts. This method shows promise for creating large corpora of known target speakers.

Figures (3)

• Figure 1: Semi-automatic pipeline proposed for the extraction of clean target speaker segments
• Figure 2: Histogram of similarity scores for different types of pairs of speakers from the INA speaker dictionary
• Figure 3: Precision and Recall over threshold values on the INA speaker dictionary