Kitten TTS: 15 Million Parameters, 25MB on Disk, Zero GPU — and It Sounds Genuinely Good

Text-to-speech has quietly become a subscription trap. The big cloud voice services charge by the character, require internet connections, and ship your users' text to someone else's servers. So when Kitten TTS showed up — 15 million parameters, 25MB on disk, running on CPU — we had to know: is small actually good enough?

Short answer: for a surprising number of uses, yes. Long answer below.

What Kitten TTS Is

Kitten TTS is an open-weight, ultra-compact text-to-speech model. The headline specs are real: the model is about 15 million parameters and roughly 25MB on disk, and it runs in real time on a plain CPU. No GPU, no API key, no per-character billing. It works on a Raspberry Pi, an old laptop, or a $5 VPS instance.

For comparison, the cloud voices most products use are billion-parameter-class models behind a metered API.

The Listening Test

We generated the same set of sentences — news copy, conversational chatbot responses, and a product FAQ — with Kitten TTS and with two mainstream cloud voices, then had five people rate them blind.

The results:

• Conversational text: Kitten scored closer to the cloud voices than we expected. Listeners described it as "clearly synthetic but pleasant" versus "clearly synthetic and annoying" for older small models.
• News/long-form: The gap widened. Cloud voices handled complex sentences and emphasis more naturally. Kitten's prosody flattened out over long passages.
• Numbers and names: This is Kitten's weakest spot. It mangled a few dollar amounts and one street name. Cloud voices nailed all of them.

Overall, listeners correctly identified Kitten as the smaller model about 80% of the time — but rated it perfectly usable for functional content.

Where It Actually Makes Sense

Based on our testing, Kitten TTS is a great fit for:

• Voice notifications and alerts — reading order statuses, server alerts, or smart-home announcements
• Privacy-sensitive applications — the text never leaves your device, which matters for healthcare, legal, and kids' apps
• Offline products — embedded devices, kiosks, aviation-mode apps
• High-volume, low-glamour use — if you're synthesizing a million characters a month, "free and local" beats "$16 per million characters" very quickly

It's a poor fit for audiobooks, marketing videos, or anything where voice is the product. There, the cloud voices still win, and it's not close.

The Math That Matters

We ran the numbers on a realistic workload: an app reading 500,000 characters of notifications per month. Cloud TTS: roughly $8–16/month, forever, plus network dependency. Kitten TTS: a one-time integration, running on hardware you already have. Over three years, that's a difference of several hundred dollars — plus zero privacy exposure.

The Verdict

Kitten TTS won't replace premium cloud voices, and it doesn't try to. What it does is make decent, private, free speech synthesis possible on hardware you already own — and that unlocks a whole category of applications that were previously uneconomical. For developers building notification-heavy or offline products, it's an easy recommendation. Everyone else, check back as the small-model race heats up, because this category is improving fast.