We did not shrink a cathedral. We packed a person.
Each seat is a 512-byte profile (256×f16) in a real .aarm.
One hundred seats live in 50,950 bytes on disk. Face, line, slot —
plug into any LLM TTS. Not a 7B. Not llama. How it packed stays packed.
This roster JSON plus a ~16 KB .aarm speaker profile per seat. Plug the slot into any LLM TTS call. The page preview is browser speech. Not a 7B. Not llama.
Everyone else ships a warehouse and calls it a voice. We ship the individual: a packed seat you can hold. The LLM you already run does the thinking. AhanaVoice is the mouth.
512-byte profile + this roster. We refuse to dump a 7B into flash and call it tiny.
Plug voice=<slot> into OpenAI-style /v1/audio/speech. Their reasoner. Our mouth. Packed.
Each seat starts in training. You approve, reject, or fix. Only approved seats are listed for sale.
Live pack jeremiah-av-experts-cpu.aarm = 50,950 bytes. One hundred f16×256 profiles = 512 bytes each. CPU serve RSS ≈ 47.5 MiB, no GPU.
| What | Bytes (measured / published) | Vs 512 B seat |
|---|---|---|
| Our speaker profile | 512 | 1× |
| Our 100-seat .aarm | 50,950 | 99.5× one seat |
| Orca complete neural TTS | 7,340,032 (7 MiB) | 14,336× |
| Piper | 63,963,136 (61 MiB) | 124,928× |
| Kokoro | 357,564,416 (341 MiB) | 698,368× |
| Cloud TTS | — | not on device |
7 MiB ÷ 512 B = 14,336. Whole pack vs one Orca engine: 7,340,032 ÷ 50,950 ≈ 144× smaller — and that pack holds 100 seats. Serve: one CPU process, ~47.5 MiB RSS, loopback, no GPU. Faces on this site are storefront art, not the download.
$14.99
One packed tiny seat in a .aarm. On-device playback.
$9.99
Two seats. Same listed bundle as LucaVoice.
$24.99
Three seats. Equal list on both storefronts.
Stripe checkout lands next. Also sold at lucavoice.com. No celebrity clones.
Different people from LucaVoice. Same prices. The slot on the card is the slot you download. Plug it into any LLM TTS. Not a celebrity clone.