Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TidyVoice: A Curated Multilingual Dataset for Speaker Verification Derived from Common Voice

Aref Farhadipour, Jan Marquenie, Srikanth Madikeri, Eleanor Chodroff

TL;DR

This work addresses the scarcity of large-scale, publicly available multilingual read-speech data for robust speaker verification by introducing TidyVoice, a tidied Common Voice-derived dataset with two evaluation partitions: Tidy-M (monolingual speakers across 81 languages) and Tidy-X (multilingual speakers across 40 languages). The authors train ResNet-34 and ResNet-293 embeddings using ArcFace loss, leveraging pre-training on VoxCeleb/VoxBlink2 and subsequent fine-tuning on Tidy-M and Tidy-X, achieving a peak $0.35\%$ $EER$ on the Tidy-M fine-tuned model and demonstrating improved generalization to out-of-domain conversational data in CANDOR. They show that fine-tuning on the large, diverse Tidy-M data yields substantial gains, with ResNet-293 typically outperforming ResNet-34, and provide detailed language-specific improvements. The dataset, evaluation trials, and trained models are publicly released to foster reproducible research in multilingual speaker verification and anti-spoofing, with potential for dynamic expansion as Mozilla Common Voice grows.

Abstract

The development of robust, multilingual speaker recognition systems is hindered by a lack of large-scale, publicly available and multilingual datasets, particularly for the read-speech style crucial for applications like anti-spoofing. To address this gap, we introduce the TidyVoice dataset derived from the Mozilla Common Voice corpus after mitigating its inherent speaker heterogeneity within the provided client IDs. TidyVoice currently contains training and test data from over 212,000 monolingual speakers (Tidy-M) and around 4,500 multilingual speakers (Tidy-X) from which we derive two distinct conditions. The Tidy-M condition contains target and non-target trials from monolingual speakers across 81 languages. The Tidy-X condition contains target and non-target trials from multilingual speakers in both same- and cross-language trials. We employ two architectures of ResNet models, achieving a 0.35% EER by fine-tuning on our comprehensive Tidy-M partition. Moreover, we show that this fine-tuning enhances the model's generalization, improving performance on unseen conversational interview data from the CANDOR corpus. The complete dataset, evaluation trials, and our models are publicly released to provide a new resource for the community.

TidyVoice: A Curated Multilingual Dataset for Speaker Verification Derived from Common Voice

TL;DR

This work addresses the scarcity of large-scale, publicly available multilingual read-speech data for robust speaker verification by introducing TidyVoice, a tidied Common Voice-derived dataset with two evaluation partitions: Tidy-M (monolingual speakers across 81 languages) and Tidy-X (multilingual speakers across 40 languages). The authors train ResNet-34 and ResNet-293 embeddings using ArcFace loss, leveraging pre-training on VoxCeleb/VoxBlink2 and subsequent fine-tuning on Tidy-M and Tidy-X, achieving a peak on the Tidy-M fine-tuned model and demonstrating improved generalization to out-of-domain conversational data in CANDOR. They show that fine-tuning on the large, diverse Tidy-M data yields substantial gains, with ResNet-293 typically outperforming ResNet-34, and provide detailed language-specific improvements. The dataset, evaluation trials, and trained models are publicly released to foster reproducible research in multilingual speaker verification and anti-spoofing, with potential for dynamic expansion as Mozilla Common Voice grows.

Abstract

The development of robust, multilingual speaker recognition systems is hindered by a lack of large-scale, publicly available and multilingual datasets, particularly for the read-speech style crucial for applications like anti-spoofing. To address this gap, we introduce the TidyVoice dataset derived from the Mozilla Common Voice corpus after mitigating its inherent speaker heterogeneity within the provided client IDs. TidyVoice currently contains training and test data from over 212,000 monolingual speakers (Tidy-M) and around 4,500 multilingual speakers (Tidy-X) from which we derive two distinct conditions. The Tidy-M condition contains target and non-target trials from monolingual speakers across 81 languages. The Tidy-X condition contains target and non-target trials from multilingual speakers in both same- and cross-language trials. We employ two architectures of ResNet models, achieving a 0.35% EER by fine-tuning on our comprehensive Tidy-M partition. Moreover, we show that this fine-tuning enhances the model's generalization, improving performance on unseen conversational interview data from the CANDOR corpus. The complete dataset, evaluation trials, and our models are publicly released to provide a new resource for the community.
Paper Structure (5 sections, 2 figures, 2 tables)

This paper contains 5 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Datasets and Conditions
  3. Experimental setup
  4. Results and Discussion
  5. Conclusion

Figures (2)

  • Figure 1: Language distribution in the train (outer circle) and test (inner circle) sets for the Tidy-M (right) and Tidy-X (left) partitions. For example, the training data for Tidy-M includes around 40K and 940 speakers for "en" and "tr" respectively.
  • Figure 2: EER as a function of training audio hours per language (log–log scale). Point size reflects the number of test speakers, and labels denote language codes. The green line shows a power-law fit, excluding languages with zero EER.