click any column header to sort

TTS Bench — Speed — 2026-05-25_0102

Rig: windows-5090 — AMD Ryzen 9 9950X3D 16-Core Processor (16C) · NVIDIA GeForce RTX 5090 32GB · 126 GB RAM · Windows 11

Label: patch · cloning · fixed + MOSS — ref chris_hemsworth_15s.wav

TTFA = time to first audio (ms; lower is better). RTFx = realtime speed (× realtime; higher is better; e.g. 10× means 10 sec of audio generated per 1 sec of compute). Cold = first run after process start; warm = subsequent runs.

Speed winners

Fastest (this cloning run): F5-TTS v1 (cuda) — 5.50× warm RTFx, 1.06s warm TTFA

Model	Device	TTFA cold	TTFA warm	RTFx cold	RTFx warm	Peak RAM	Peak VRAM	Size	Released
F5-TTS v1	cuda	1.56s	1.06s	3.75×	5.50×	2.67 GB	883 MB	330M	Oct 2024
MOSS-TTS-Nano	cuda	5.59s	4.87s	1.34×	1.66×	2.42 GB	910 MB	100M	Apr 2026
IndexTTS-2	cpu	5.77s	4.87s	1.20×	1.47×	5.54 GB	—	1.5B	Jun 2025
MOSS-TTS v1.0	cuda	5.44s	4.73s	1.31×	1.44×	2.10 GB	22.83 GB	8B	Feb 2026
IndexTTS-2	cuda	6.80s	5.66s	1.03×	1.26×	4.05 GB	7.71 GB	1.5B	Jun 2025
MOSS-TTS-Nano	cpu	5.69s	5.04s	1.16×	1.25×	3.05 GB	—	100M	Apr 2026
Dia 1.6B-0626	cuda	13.79s	11.45s	0.37×	0.50×	4.42 GB	4.82 GB	1.6B	Jun 2025
F5-TTS v1	cpu	71.35s	71.18s	0.08×	0.08×	2.71 GB	—	330M	Oct 2024