KittenTTS
A tiny, open-source text-to-speech library (as small as 25MB) that produces natural-sounding voices on ordinary CPUs — no GPU required.
🔗 Visit KittenTTS
Description
Good text-to-speech usually means either a cloud API that costs money per character, or a local model so large it needs an expensive GPU to run in real time. KittenTTS is built to avoid both: it's a small enough model — as little as 25MB — that it generates natural speech on a regular CPU, making it practical to embed directly into mobile apps, edge devices, or any service where a GPU isn't an option.
KittenTTS is an Apache 2.0-licensed, ONNX-based text-to-speech library optimized for CPU inference, with models ranging from 15M to 80M parameters (25-80MB on disk). It ships with 8 built-in voices (Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, Leo), 24kHz output, adjustable speech rate, and text normalization for abbreviations, dates, and other special formatting. With 15,200+ GitHub stars, it's one of the more popular lightweight TTS projects, and commercial support plus custom voice development are available for teams that need it.
💬 Our review
The short version: if you need to add a voice to an app without paying per-character cloud fees or requiring a GPU, KittenTTS's small model size and CPU-only inference make it one of the more practical open-source choices available right now.
Against cloud TTS services (ElevenLabs, Google Cloud TTS, Amazon Polly), KittenTTS trades some voice naturalness and variety — 8 built-in voices versus dozens on paid clones-and-cloning platforms — for zero per-use cost and full offline operation. Against other open-source TTS projects that need a GPU for real-time synthesis, KittenTTS's CPU-only design is the real differentiator, since it removes the biggest deployment obstacle for edge and mobile use cases. The 15,200+ stars suggest real adoption, and the availability of commercial support is a reassuring sign for teams nervous about depending on an open-source project without a company behind it — though voice quality at this model size still won't match the best cloud voice-cloning services for consumer-facing polish.
💰 Pricing
📊 Global score
🤖 AI-enriched data
100% open source (Apache 2.0) ; support commercial et voix sur-mesure disponibles en option payante
Pros
Modèle minuscule (25-80 Mo), fonctionne sans GPU
Gratuit, open source, aucun coût par caractère
8 voix intégrées, sortie 24 kHz
Normalisation de texte intégrée (dates, abréviations)
Cons
Moins de voix/variété que les services cloud payants
Qualité vocale en retrait face aux meilleurs clones IA du marché
Support commercial payant en option seulement