Research map / Image generation / Autoregressive and latent generation
Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale
Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, Wei-Ning Hsu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use optimal-transport path instead of diffusion path (flow matching) (ablation) | Voicebox audio model · Zero-shot TTS 50K updates | WER: 76.0 -> 2.5 | Improved |
| use flow matching OT rather than diffusion path at low NFE (ablation) | Voicebox audio model · Zero-shot TTS NFE=4 | WER: 11.5 -> 2.4 | Improved |
| use flow matching instead of score matching with diffusion path (ablation) | Voicebox audio model · Zero-shot TTS 150K updates | WER: 5.1 -> 2.1 | Improved |
| use flow matching instead of regression objective (ablation) | Voicebox audio model · Zero-shot TTS cross-sentence | SIM-r: 0.520 -> 0.597 | Improved |
| use flow matching instead of regression objective (ablation) | Voicebox audio model · Diverse sampling | FSD: 278.8 -> 242.5 | Improved |
| use flow matching instead of regression objective (ablation) | Voicebox audio model · Zero-shot TTS cross-sentence | WER: 2.0 -> 2.1 | No clear effect |
| compute loss only on masked frames (ablation) | Voicebox audio model · Zero-shot TTS cross-sentence | SIM-r: 0.528 -> 0.597 | Improved |
| compute loss only on masked frames (ablation) | Voicebox audio model · Zero-shot TTS cross-sentence | WER: 2.0 -> 2.1 | No clear effect |
| scale training data from 60 to 60,000 hours (ablation) | Voicebox · Zero-shot TTS | SIM-r: 0.151 -> 0.645 | Improved |
| scale training data from 6,000 to 60,000 hours (ablation) | Voicebox · Diverse sampling | FSD: 195.52 -> 214.38 | Hurt |
| use flow-matching duration model instead of regression duration model (ablation) | Voicebox VB-En · ASR trained on synthetic data, test-clean | WER: 7.1 -> 3.1 | Improved |
| use flow-matching duration model instead of regression (ablation) | Voicebox VB-En · Diverse speech generation | WER: 3.1 -> 5.6 | Hurt |
Closest papers by meaning
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.89)
- Vevo: Controllable Zero-Shot Voice Imitation with Self-Supervised Disentanglement (ICLR 2025 · similarity 0.89)
- Metis: A Foundation Speech Generation Model with Masked Generative Pre-training (NeurIPS 2025 · similarity 0.88)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.88)
- AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation (NeurIPS 2026 · similarity 0.88)
- RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
