Abstract Removing intermediate representations and separately trained decoding stages has become an important direction in generative modeling. In text-to-speech, however, high-quality systems are still commonly built through an intermediate acoustic representation before waveform synthesis. In this work, we present BareWave, a fully waveform-native framework for direct text-to-wave generation in flow-matching TTS. We consider this setting to raise three training challenges: raw-waveform modeling lacks a strong pretrained representational scaffold, there is no single noise schedule that remains optimal throughout training, and data-space perceptual objectives do not automatically share the temporal structure of the velocity-space flow objective. As a result, direct waveform training is hard to optimize efficiently, hard to push toward a strong final operating point with a fixed recipe, and hard to integrate effective perceptual refinement. Guided by this view, we develop a direct text-to-wave training framework that combines training-time representation alignment, staged noise scheduling, and velocity-aware perceptual alignment (VAPA), while preserving a single waveform-native inference path without pretrained components at test time. Experiments on zero-shot voice cloning show that strong intelligibility, speaker similarity, and naturalness can be achieved under a fully waveform-native inference path, supporting waveform-native flow-matching TTS as a practical direction.

Contents

• Model Overview
• Audio Samples

This page is for research demonstration purposes only.

Model Overview

Figure 1: Comparison with Existing Methods

Mainstream TTS pipelines rely on intermediate representations and separate waveform decoders, whereas our framework is fully waveform-native.

Figure 2: Method Overview

Overview of our training framework combining representation alignment, staged noise scheduling, and velocity-aware perceptual alignment.

Main Contributions

• We present BareWave, a fully waveform-native framework for zero-shot TTS, with no intermediate acoustic representation, pretrained inference-time component, or separate vocoder.
• We develop a direct text-to-wave training recipe that combines representation alignment, staged noise scheduling, and velocity-aware perceptual alignment while preserving a single waveform-native inference path.
• We observe that data-space perceptual losses are implicitly down-weighted relative to the velocity-space flow objective near cleaner timesteps, and propose Velocity-Aware Perceptual Alignment (VAPA) to match the temporal structure of the flow objective.

Audio Samples

These samples correspond to the systems compared in Table 2 of the paper under the zero-shot voice cloning setting. Both our waveform-native methods and the F5-TTS baseline are trained on the same Emilia English subset, and all samples here are drawn from LibriSpeech-PC test-clean.

Prompt is the reference audio for zero-shot cloning, and Ground Truth is the original target speech. F5-TTS (mel-based) is the strong intermediate-representation baseline. Simple direct-wave baseline is the simplest direct-wave baseline, which predicts waveform patches directly from raw waveform patch prompts. Ours (basic training) uses the same waveform-native architecture as our method, but is trained only with the basic MSE flow-matching loss under a fixed logit-normal noise schedule, without the proposed training design. Ours (proposed training scheme) is our full waveform-native system with representation alignment, staged noise scheduling, and velocity-aware perceptual alignment.