Research map / Architectures and efficiency / RNNs and normalization
TVTSyn: Content-Synchronous Time-Varying Timbre for Streaming Voice Conversion and Anonymization
Waris Quamer, Mu-Ruei Tseng, Ghady Nasrallah, Ricardo Gutierrez-Osuna
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 |
|---|---|---|---|
| add Global Timbre Memory (ablation) | TVTSyn · TVT block ablation (voice conversion) | NISQA MOS: 3.45 -> 3.91 | Improved |
| add learnable prior prototypes in GTM (ablation) | TVTSyn · TVT block ablation (voice conversion) | NISQA MOS: 3.62 -> 3.91 | Improved |
| use slerp instead of linear interpolation (ablation) | TVTSyn · TVT block ablation (voice conversion) | NISQA MOS: 3.75 -> 3.91 | Improved |
| add learned gating instead of fixed alpha=0.5 (ablation) | TVTSyn · TVT block ablation (voice conversion) | NISQA MOS: 3.80 -> 3.91 | Improved |
| use 48 GTM tokens instead of 24 (ablation) | TVTSyn · TVT block ablation (voice conversion) | NISQA MOS: 3.81 -> 3.91 | Improved |
| use 48 GTM tokens instead of 12 (ablation) | TVTSyn · TVT block ablation (voice conversion) | NISQA MOS: 3.73 -> 3.91 | Improved |
| add TVT processing block (ablation) | TVTSyn · Voice conversion (objective) | NISQA-MOS: 3.44 -> 3.91 | Improved |
| add factorized VQ bottleneck (ablation) | TVTSyn · Voice conversion (objective) | NISQA-MOS: 3.42 -> 3.91 | Improved |
| add both TVT and VQ (ablation) | TVTSyn · Voice conversion (objective) | NISQA-MOS: 3.179 -> 3.91 | Improved |
| add TVT or VQ (ablation) | TVTSyn · Voice conversion (objective) | Src-SIM (anonymization): 0.48 -> 0.47 | No clear effect |
| add TVT block components (GTM, prior, slerp, gate) (ablation) | TVTSyn · TVT block ablation (voice conversion) | Src-SIM: 0.48 -> 0.47 | No clear effect |
| TVTSyn versus SLT24 on streaming anonymization | TVTSyn · VPC'24 speaker anonymization (LibriSpeech) | WER (%): 5.70 -> 5.35 | Improved |
Closest papers by meaning
- Do Not Mimic My Voice : Speaker Identity Unlearning for Zero-Shot Text-to-Speech (ICML 2025 · similarity 0.84)
- StableVC: Style Controllable Zero-Shot Voice Conversion with Conditional Flow Matching (AAAI 2025 · similarity 0.84)
- Vevo: Controllable Zero-Shot Voice Imitation with Self-Supervised Disentanglement (ICLR 2025 · similarity 0.84)
- CSSinger: End-to-End Chunkwise Streaming Singing Voice Synthesis System Based on Conditional Variational Autoencoder (AAAI 2025 · similarity 0.84)
- BinauralFlow: A Causal and Streamable Approach for High-Quality Binaural Speech Synthesis with Flow Matching Models (ICML 2025 · similarity 0.84)
- SelfVC: Voice Conversion With Iterative Refinement using Self Transformations (ICML 2024 · similarity 0.83)
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.
