One-Step Text-to-Audio Generation
MeanFlow-anchored Fréchet-distance post-training. Listen to FdAudio against MeanAudio-S-Full, ConsistencyTTA, AudioLCM, and the ground-truth recordings on AudioCaps prompts.
Single-step NFE = 1
Only 120M parameters
1.79 s / 8 clips (V100)
SOTA one-step on AudioCaps
| # | Prompt | FdAudio ours NFE = 1 | MeanAudio-S-Full NFE = 1 | ConsistencyTTA NFE = 1 | AudioLCM NFE = 2 | Ground-Truth recording |
|---|---|---|---|---|---|---|
| 01 | Water is falling, splashing and gurgling | |||||
| 02 | A door closes followed by a toilet flushing | |||||
| 03 | An engine idling with light wind | |||||
| 04 | A series of bell chime | |||||
| 05 | A vehicle engine revving then accelerating at a high rate as a metal surface is whipped followed by tires skidding | |||||
| 06 | A vehicle engine starting up then running idle | |||||
| 07 | A muffled helicopter engine flying | |||||
| 08 | Ocean waves crashing as wind blows into a microphone | |||||
| 09 | An insect buzzing as plastic clacks and plastic slaps a hard surface | |||||
| 10 | A sewing machine operating during several metal clacks |
FdAudio · MeanFlow-Anchored Fréchet-Distance Post-Training for One-Step Text-to-Audio Generation