Pocket TTS self-hosted vs hosted TTS: a break-even formula
Kyutai Pocket TTS is MIT-licensed but needs a server. Cost per audio hour is server price divided by speed. On Sume it is $0.0475 per 1,000 characters.
Self-hosting Pocket TTS costs server price per hour / real-time factor per hour of audio, and hosted Sume TTS costs characters per audio hour / 1,000 * $0.0475. With a server at $0.50 an hour running at 2.3x real time, one audio hour costs $0.22 self-hosted. Sume costs $2.565 for the same hour only if your script runs 900 characters per audio minute, an input to replace with your own number. The gap is real. So are the things the formula leaves out.
The inputs from the vendor
Kyutai's GitHub page lists Pocket TTS as a 100M-parameter model under the MIT license, covering English, French, German, Portuguese, Italian, Spanish and Dutch. It reports about 6x real time on a MacBook Air M4 and about 2.3 to 2.5x on x86 cloud VMs, with roughly 200 ms to the first chunk (Kyutai on GitHub, read 2026-10-04). It needs PyTorch 2.5 or newer, supports voice cloning from a WAV, and does not support silences or pauses in the text.
Cost per audio hour
Server cost divided by 2.3, the lower end of Kyutai's x86 range. This assumes one stream per machine and a machine that is busy the whole hour.
| Server $ per hour | Pocket TTS $ per audio hour | Sume at 900 characters per minute |
|---|---|---|
| 0.25 | 0.109 | 2.565 |
| 0.50 | 0.217 | 2.565 |
| 1.00 | 0.435 | 2.565 |
| 2.00 | 0.870 | 2.565 |
What the formula leaves out
- Utilization. A server that sits idle 90 percent of the day makes the self-hosted hour ten times dearer.
- Languages and voices. Pocket TTS covers seven languages, and you handle voice rights yourself.
- Pauses. The model reads no silence markers, so a slow reading needs another approach.
- Everything after the voice. Sume's detach, join and render steps ($0.01, $0.01 and $0.10 per minute) are separate jobs you would otherwise run yourself.
How to decide
Measure your own characters per audio minute from one Sume take (the measuring guide), put your real server price in the table, and compare. If you voice more than a few hours a week and have an engineer to run it, self-hosting can pay. If you voice a few minutes per video, Sume's $0.0475 per 1,000 characters is the lower-effort path, and Microsoft's hosted MAI-Voice-2.1 at $22 per million characters is a third option (Microsoft AI, read 2026-10-04).
Sources
Related posts
More in Comparisons
- One podcast episode's audio bill: MAI, Lyria and transcript vs Sume
A 5,000-character intro, one music track and a 60-minute transcript cost $0.695 on MAI Flash, Lyria 3.5 and MAI-Transcribe, and $0.9625 on Sume.
- LiveTranslate 2.3 s latency vs STT plus TTS
Qwen3.8-LiveTranslate cut average lagging from 2.8 to 2.3 seconds. Sume has no live interpreter, but stt_create and tts_create can build an offline dub.
- Recast, Omni edit, Genjutsu or a new shot: which ad variant method?
Four ways to make an ad variant on Sume, limits side by side: Recast swaps people, Omni edit changes by prompt, Genjutsu moves motion, a new shot starts over.
- Recast one holiday ad for four markets: fal vs Sume cost
Four 15-second Recast runs cost $18.00 at fal's $0.30 a second and $22.50 on Sume at $0.375 a second, both at 768p. See the math and what the gap buys.
Written by Sume