Kokoro

Kokoro-82M is an 82-million-parameter open-weight text-to-speech model by hexgrad, Apache 2.0 licensed. Despite being a fraction of the size of competitors (467M–1.2B), it ranks #2 in TTS Arena behind only ElevenLabs, produces natural 24kHz audio, and ships 54 voice presets across 9 languages (English US/UK, Japanese, Spanish, French, Hindi, Italian, Portuguese, Mandarin) with voice blending. The ONNX build runs at 3.3x real-time on an iPhone 13 Pro — no GPU required — and API serving costs under $1 per million characters.

The honest picture: 82M parameters buys efficiency, not the full expressive range of larger dialogue models — no multi-speaker conversation mode, no nonverbal vocalizations, and voice cloning is not its strength (it uses fixed voice presets rather than zero-shot cloning). Grapheme-to-phoneme quality depends on the wrapper (misaki for English is best). For narration, audiobooks, and voiceovers it is the efficiency king; for cloning a specific voice, look at Dia or Chatterbox instead.