AI
NVAlign: Direct-Gradient Optimization for Non-Verbal Control in Continuous Autoregressive Flow Matching Text-to-Speech
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 24/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 05:26:01 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 23:18:24 UTC.
Briefings
NVAlign is a direct-gradient post-training framework for non-verbal vocalization (NVV) in continuous autoregressive flow-matching text-to-speech models. The method uses supervised fine-tuning on NVV-annotated data followed by a two-step gradient surrogate to backpropagate rewards through the flow sampler, updating both the autoregressive backbone and acoustic flow head. Results on NVV-SuperBench and human evaluations show NVAlign significantly improves tag-following accuracy compared to SFT and Flow-GRPO baselines. Audio samples are available at https://nvalign.github.io/.
Key points
- NVAlign uses a direct-gradient post-training framework to improve non-verbal vocalization (NVV) control in continuous autoregressive flow-matching text-to-speech models.
- The framework employs a two-step gradient surrogate to backpropagate rewards through the flow sampler, jointly updating the autoregressive backbone and acoustic flow head.
Context and limitations
Results are limited to NVV-SuperBench and human listening evaluations; specific quantitative metrics for NVV-SuperBench are not provided in the evidence.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 05:26:01 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibm7ajgn35l7ecv4q4jlzt5cu5642gf75urcdz4fbichhenuq663i