Abstract
This paper introduces GLOBE, a high-quality English corpus featuring worldwide accents, specifically designed to address the limitations of current zero-shot speaker adaptive Text-to-Speech (TTS) systems. While voice cloning technology has advanced, existing models often exhibit poor generalizability when adapting to speakers with diverse accents. GLOBE addresses this by providing utterances from 23,519 speakers across 164 global accents, accompanied by detailed metadata. By rigorously filtering and enhancing the massive but noisy Common Voice dataset, GLOBE offers 535 hours of clean speech data at a 24 kHz sampling rate. Benchmark results demonstrate that models trained on GLOBE achieve superior speaker similarity and comparable naturalness to those trained on standard, less diverse corpora.
Introduction
Zero-shot speaker adaptive TTS, commonly known as voice cloning, aims to synthesize a target speaker's voice using only a brief reference audio clip. Despite its potential, a significant "accent gap" persists. Most popular training datasets, such as LibriTTS or VCTK, are heavily biased toward US and UK English, featuring a limited number of accents. Consequently, when these models encounter speakers from regions like Singapore, India, or Nigeria, performance drops; the synthesized voice often loses its characteristic accent or sounds unnatural.
While the Mozilla Common Voice project offers the necessary accent diversity, it is traditionally unsuitable for high-quality TTS training due to background noise, low-bandwidth recordings, and missing speaker metadata. GLOBE bridges this gap by refining Common Voice into a professional-grade research corpus.
The GLOBE Pipeline
To transform raw Common Voice data into the refined GLOBE corpus, the researchers developed a multi-stage processing pipeline:
- Rigorous Filtering: Audio clips were discarded if they possessed a Signal-to-Noise Ratio (SNR) below 0 dB or a signal bandwidth below 12 kHz. Clips with internal silence exceeding 930ms—often indicative of speaker hesitation—were also removed.
- Text Re-alignment: To ensure transcript accuracy, OpenAI’s Whisper was used to re-transcribe the audio. The team calculated the edit distance between the original text and the Whisper transcript, discarding clips with significant discrepancies.
- Metadata Population: Approximately half of Common Voice speakers lack metadata. The team utilized HuBERT and EPACA-TDNN architectures to build a prediction model that filled in missing age, gender, and accent labels. This model achieved high accuracy, reaching 99.95% for gender prediction.
- Speech Enhancement: Finally, Voice Activity Detection (VAD) was applied to trim leading and trailing silences, and speech enhancement tools were used to further suppress residual background noise.
Dataset Comparison
GLOBE stands out when compared to existing industry standards:
- GLOBE: 535 hours, 23,519 speakers, 164 accents.
- Common Voice (Source): 3,347 hours, but significantly lower audio quality.
- LibriTTS / LibriTTS-R: High quality, but lacks accent diversity.
- VCTK: High quality, but limited to only 11 accents.
To quantify this diversity, the authors employed the Vendi Score (SEVS). The results proved that GLOBE maintains the high diversity of Common Voice while matching the audio quality of premium datasets like LibriTTS-R.
Key Insights and Results
Quality vs. Quantity
A critical finding of this research is the trade-off between data volume and data purity. The researchers discarded nearly 85% of the original Common Voice data to create GLOBE. For TTS applications, a smaller volume of "clean" data is far superior to a massive volume of "noisy" data.
Diversity as a Generalization Engine
The study suggests that training on 164 accents does not merely improve performance for those specific accents; it enhances the model's fundamental ability to adapt to any unseen accent. This makes GLOBE a powerful tool for building robust, globally-aware AI systems.
Performance Metrics
- Naturalness: In Mean Opinion Score (MOS) tests, GLOBE achieved a score of 4.25, effectively tying with the gold-standard LibriTTS-R (4.27) and significantly outperforming raw Common Voice (3.54).
- Speaker Similarity: When tested on accented speakers, a TTS model (YourTTS) trained on GLOBE showed significantly higher speaker similarity scores than the same model trained on LibriTTS.
- Robustness: Models trained on GLOBE exhibited the smallest performance decline when moving from standard English to diverse global accents, confirming its superior generalization capabilities.
Conclusion and Future Work
GLOBE represents a major step forward in addressing demographic bias in AI speech synthesis. While currently limited to English, the pipeline developed here could be applied to create high-quality, multi-accented corpora for other global languages. Future work will explore using the rich metadata (age, gender, accent) to build even more personalized TTS models that can further mitigate biases in speech technology.
Summary updated