Voice · hexgrad (open source)
Kokoro
A small open-source voice model, private and by far the cheapest. Sub Rosa uses it by default for narration.
Private Low costWhat sets it apart
Kokoro is an 82-million-parameter open-weights model that runs on modest hardware, which is why it is the cheapest voice in the catalog and can run in private mode. It trades some expressiveness for speed and cost.
- 82M parameters, Apache 2.0
- 54 voices, 8 languages in v1.0
- Trained for about $1,000 of GPU time
Recommended for: read a note or narrate a film.
Strengths and limits
Where it shines
- Cheapest voice in the catalog by a wide margin
- Private mode: nothing is kept
- 54 voices across 8 languages
Think twice when
- Only one French voice (ff_siwis)
- Less expressive than the larger models
- 54 voices
- Up to 5,000 characters per request in Sub Rosa
Use cases
Long article read aloud
Read this blog post aloud in a calm, neutral voice so I can listen while I cook.
A very low cost per character keeps long texts affordable.
Private narration of sensitive notes
Read my meeting summary aloud.
Private mode means nothing is kept.
Against its rivals
- Orpheus TTSOrpheus is more expressive in English but costs more and speaks English only.
- Qwen3 TTSQwen3 TTS covers 10 languages with cloning, at several times Kokoro's price.
Version history
Each version, dated, with what it changed over the one before.
- Kokoro v0.19Earlier, not in Sub Rosa
- Initial release: English only, 10 voices, trained on under 100 hours of audio
- Kokoro v1.0Latest
- Multilingual: 8 languages instead of 1
- 54 voices instead of 10, trained on a few hundred hours of audio
Benchmarks
How to read these scores: what a benchmark tells you.
Artificial Analysis Speech Arena: #10 of 10 models measured in the catalog. People listen to the same sentence read by two voices, without knowing the models, and pick the more natural one. An Elo rating: higher sounds more natural to more people. 100 points apart means the higher-rated model wins about 64% of duels.
- 1. ElevenLabs TTS v4 TurboElevenLabs1,330
- 2. ElevenLabs TTS v4ElevenLabs1,329
- 3. Gemini 3.1 Flash TTSGemini 3.1 Flash TTS1,209
- 4. ElevenLabs TTS v3ElevenLabs1,178
- 5. Gradium TTSGradium TTS1,151
- 6. MiniMax Speech-02 HDMiniMax Speech-02 HD1,114
- 7. Chatterbox HD (Resemble AI)Chatterbox HD1,102
- 8. ElevenLabs Turbo v2.5ElevenLabs1,097
- 9. ElevenLabs Multilingual v2ElevenLabs1,095
- 10. Kokoro Text to SpeechKokoro1,064
See the numbers
| Model | Elo |
|---|---|
| 1. ElevenLabs TTS v4 Turbo | 1,330 |
| 2. ElevenLabs TTS v4 | 1,329 |
| 3. Gemini 3.1 Flash TTS | 1,209 |
| 4. ElevenLabs TTS v3 | 1,178 |
| 5. Gradium TTS | 1,151 |
| 6. MiniMax Speech-02 HD | 1,114 |
| 7. Chatterbox HD (Resemble AI) | 1,102 |
| 8. ElevenLabs Turbo v2.5 | 1,097 |
| 9. ElevenLabs Multilingual v2 | 1,095 |
| 10. Kokoro Text to Speech | 1,064 |
The model in this family
| Model | Use | Privacy | Price |
|---|---|---|---|
Kokoro Text to Speechtts-kokoro | Voice | Private | $0.00157 per 1,000 characters$0.00472 a page read aloud |
Specifications
| Specification | Kokoro v1.0 |
|---|---|
| Released | January 27, 2025 |
| Parameters | 82M |
| Weights | Open, Apache 2.0 |
| Accepts | text |
| Produces | audio |
| Languages | 8 languages, 54 voices |
| Price | $0.00157 per 1,000 characters |
| Privacy | Private |
Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.
Sources
Everything on this page comes from these pages, read on October 16, 2026 or as dated on each score.