hexgrad/Kokoro-82M

Kokoro‑82M is a compact and efficient TTS model built by hexgrad, offering impressive audio quality despite its small size. The model uses a hybrid architecture combining StyleTTS 2 and ISTFTNet, enabling it to produce natural-sounding speech without requiring massive compute resources.

Licensed under Apache 2.0, Kokoro‑82M is fully open-weight, making it accessible for both personal and commercial use.

According to its GitHub README, inference is very cost-efficient — roughly less than $0.06 per hour of generated audio, or under $1 per million input characters.

Despite having only 82M parameters, Kokoro‑82M supports 54 distinct voice profiles across 8 languages, demonstrating a strong capacity for multilingual and multi-voice synthesis.

It was trained on a relatively modest dataset (under 100 hours of permissively licensed audio) and with very controlled compute costs (reported around $1,000 of A100 GPU time).

Deployment options are flexible. There's a Python inference library (kokoro), and you can run it locally (even on CPU) or deploy via API.

Voxta’s documentation also provides configuration options, such as voice selection (male/female), language codes, and device settings.

Users on Reddit praise its speed and efficiency, noting that it runs very fast even on machines with modest hardware (e.g., M1 Pro).

Others have extended its capability with batch processing, enabling high-throughput generation.

💰 Pricing / Cost Overview

Open‑weight model: The model weights are freely available under Apache 2.0, which means no licensing cost for usage.

Inference running cost: According to the README, it costs under $1 per million input characters, and about $0.06 per hour of audio when using hosted or API-based inference.

API options: TogetherAI supports Kokoro‑82M, but pricing depends on usage.

⭐ Pros & Cons

✅ Pros

Very lightweight: Only 82M parameters, making it efficient to run locally or on lower‑cost hardware.

High-quality audio: Despite small size, delivers quality comparable to much larger TTS models.

Multi-voice & multilingual: Supports 8 languages and 54 voice types.

Open-license: Apache 2.0 license allows commercial use and deployment.

Good resource efficiency: Training cost was very low; inference cost is modest.

Fast inference: Known to run quickly even on modest hardware, according to community reports.

❌ Cons

Limited training data: Only ~100 hours of audio used, which might limit expressivity vs very large models.

Voice cloning: Not designed primarily for cloning — may not perform as well as specialized cloning models. > “take your voice cloning performance expectations … divide by 10”.

Batch support: Originally lacked batch processing (though the community added it).

Limited context: With small model size, may struggle on very long or highly complex text.

🎯 Recommendation

Best for: Developers, creators, and small teams who need a fast, efficient, and open-source TTS engine for voiceovers, chatbots, accessibility, or lightweight production.

Great for: Running TTS locally, deploying on resource-constrained devices, or building cost-efficient voice applications.

Less ideal for: Projects requiring ultra-high fidelity voice cloning, or very large-scale commercial TTS with massive data and compute backing.