BareWave: waveform-native flow matching TTS eliminates intermediate representations

BareWave, presented by Wei Fan and colleagues, is a fully waveform-native text-to-speech framework that removes the traditional intermediate acoustic representation (e.g., mel-spectrograms) and separately trained decoder stage. Instead, it uses flow matching to directly map text to raw waveform. The authors identify three key challenges native to this approach: raw-waveform modeling lacks a strong pretrained representational scaffold, different training stages require different noise schedules, and data-space perceptual objectives do not share the temporal structure of the velocity-space flow objective. To address these, BareWave introduces training-time representation alignment, staged noise scheduling, and a novel velocity-aware perceptual alignment (VAPA) loss. The system preserves a single waveform-native inference path with no pretrained components at test time.

Experiments on zero-shot voice cloning demonstrate that BareWave matches or surpasses traditional cascaded systems in intelligibility (measured by CER), speaker similarity (speaker embedding cosine similarity), and naturalness (MOS). The model achieves strong results without relying on externally trained feature extractors or separate vocoders. This work shows that fully waveform-native flow-matching TTS is not only feasible but practical, potentially simplifying TTS pipelines and enabling more direct control over voice characteristics. The project page includes audio demos for subjective evaluation.