📄 Showing the "Simple" version — "Advanced" is not ready yet. Add it to favorites to help prioritize it. Ask for the full read

Narrating a long text, like an audiobook, used to require the computer to keep a detailed picture of every split second of sound in mind, as if for an hour-long conversation you had to draw a millimeter-accurate map of the entire route. These 'maps' were enormous, and even powerful machines choked, assembling speech piece by piece.

The ZipL-Dialog method takes a different approach: it first compresses the sound into a compact blueprint, like a coiled spring storing potential energy. In this compact realm, with less data, the spring 'uncoils' into finished speech. The trainable system learns to compress and restore sound without sacrificing expressiveness, and the ZipFormer model operates directly within the compressed representation.

The result: peak memory consumption plummeted by a factor of 11.22, and generation speed jumped 2.23 times. Now an hour-long voice recording flows without pauses, as if a filing cabinet full of disks were replaced by a single flash drive.

A remarkable twist: in compressed form, sound becomes so compact that an entire novel could fit in a phone's RAM, and synthesis begins to feel like playing an invisible instrument, where each note is plucked from mathematical tautness.

Original abstract

as written by the authors arXiv:2607.12496

Zero-shot dialog TTS benefits from flow-matching, but minute-scale generation on dense mel-spectrograms causes severe memory bottlenecks, often forcing unnatural chunked synthesis. We propose ZipL-Dialog, which shifts conditional flow-matching into a 4x time-compressed (25 Hz) latent space. To preserve acoustic fidelity under compression, we employ a deterministic mel autoencoder with auxiliary mel-domain supervision and optimize the ZipFormer's hierarchical downsampling schedule. Experiments show that ZipL-Dialog reduces maximum peak GPU memory by 11.22x and accelerates inference by 2.23x over the baseline, substantially lowering the memory footprint of single-pass multi-minute dialog synthesis while maintaining perceptual naturalness.