VoxMel TTS: tokenizer-free autoregressive speech synthesis via Haar-factorized mel
VoxDance-Lab · hf · 2026-11-17
VoxDance Lab released the VoxMel TTS technical report, a tokenizer-free autoregressive text-to-speech framework.
- Core idea: an invertible two-axis Haar transform decomposes mel spectrograms into a Coarse-Mel tier and residual tiers
- The autoregressive stage predicts next coarse frames only (LM backbone + Flow Matching Head, no learned acoustic tokenizer); a non-autoregressive stage reconstructs temporal and frequency residuals, with a Vocos vocoder producing 24 kHz audio
- Conditioning uses target text (Qwen2.5 tokens), reference transcript, and reference speech
- Reported stable high-fidelity generation with strong recognition accuracy and speaker similarity
Skipping discrete acoustic tokens avoids their information loss — an interesting architecture direction for high-fidelity TTS.
More from Multimodal
- Seedance 2.5 demo: a wholesome 3D cartoon clip with the full prompt shared — aftahi_ai · 2026-11-17
- AI-generated images capture rain falling from a sunny sky with stunning light — Kyrannio · 2026-11-17
- Carbon model nails resin-printable chess pieces on the first pass — Scobleizer · 2026-11-17
- Retro sci-fi AI image set "The Last Horizon" dazzles with crystal cities — Kyrannio · 2026-11-17
- Seedance 2.5 video is so realistic viewers mistake it for real footage — SimplyAnnisa · 2026-11-17
- Open-source ComfyUI-Prism Pipeline integration brings AI generation into VFX workflows — Rico571 · 2026-11-17