Text-to-SpeechPrivate

Kokoro Text to Speech

An ultra-lightweight, open-weight text-to-speech model delivering studio-quality synthesis with incredible speed and efficiency.

Maker
hexgrad
Modality
Speech
License
Apache-2.0
Open weights
Yes (Apache-2.0)

Overview

What is Kokoro Text to Speech

Kokoro is an open-weight, ultra-efficient text-to-speech (TTS) model developed by hexgrad. With only 82 million parameters, it delivers highly natural, studio-quality audio synthesis that rivals models ten times its size, while running significantly faster and more cost-effectively on Venice's private platform.

Running it privately on Venice

On Venice, you can access Kokoro with zero retention, ensuring your synthesized text is never stored or profiled. By routing requests privately, Venice bypasses the surveillance pipelines of Big Tech, giving you permissionless, sovereign, and uncensored access to high-fidelity audio generation. This allows you to integrate open-source TTS into your workflows without sacrificing data privacy.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Incredible efficiency: At just 82M parameters, it is highly optimized, generating audio up to 210x real-time on high-end GPUs.
  • Studio-quality naturalness: Outperformed much larger models like XTTS v2 and MetaVoice in TTS Arena ELO ratings despite its tiny footprint.
  • Permissive Apache-2.0 license: Fully open-source and open-weights, allowing unrestricted commercial deployment and self-hosting.
  • Broad voice library: Ships with 54 high-quality preset voices across multiple accents and languages.
  • Extremely low latency: Ideal for real-time applications, interactive agents, and cost-sensitive pipelines.
Limitations
  • No native zero-shot voice cloning: Unlike some massive closed models, it relies on its preset voice library rather than instant cloning from a reference audio clip.
  • Short-form optimization: Best suited for chunked text; extremely long, unstructured paragraphs may require preprocessing or pipeline chunking for optimal flow.
  • Third-party hosting: Served via a third-party provider on Venice, meaning anonymized requests are forwarded rather than run on native Venice hardware.

Samples

Sample outputs

Generated on Venice with our standard prompt suite — the same scripts we run through every model of this type, so you can judge it like-for-like.

Narration

On Venice, your prompts are processed privately and never stored, profiled, or used to train anyone's model.

Conversational

Wait — so I can run a private voice model with zero data retention, and pay only for what I use? That's genuinely useful.

Expressive range

Three… two… one… liftoff! The rocket roared into the night sky as the crowd erupted in cheers.

Compare every speech model on these scripts

Specifications

Datasheet

Maker
hexgrad
Released
January 27, 2025
Architecture
Flow matching + Vocos vocoder (StyleTTS2-based)
Parameters
82 Million
Preset Voices
54 voices
Languages
Multilingual (English, Spanish, French, Japanese, and more)
Open weights
Yes (Apache-2.0)
Privacy on Venice
Private — zero retention
Available on Venice since
Mar 2025

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/audio/speech \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tts-kokoro",
    "input": "On Venice, your prompts are processed privately.",
    "voice": "af_sky",
    "response_format": "mp3"
  }' --output speech.mp3

Pricing

What it costs on Venice

Billed per character on Venice: $3.50 per 1M characters of synthesized speech.

Characters / 1M
$3.50
Per 1M characters

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelBest forParametersOpen weightsPrice (Venice)Primary Strength
Kokoro Text to SpeechThe undisputed champion of lightweight, open-weight TTS.82MYes$3.50 / 1M charsExtreme speed & efficiency
Chatterbox HD (Resemble AI)Excellent for custom voice cloning, but significantly more expensive.ProprietaryNo$50 / 1M charsHigh-fidelity voice cloning
ElevenLabs Turbo v2.5Industry benchmark for expressive speech, but closed-source and costly.ProprietaryNo$62.50 / 1M charsUltra-realistic expression
Orpheus TTSAnother open-weights option, though Kokoro remains far more cost-effective.UnknownYes$62.50 / 1M charsOpen-weights synthesis

The undisputed champion of lightweight, open-weight TTS.

Use cases

What it is good for

  1. 01Real-time conversational AI assistants and voice agents requiring ultra-low latency.
  2. 02E-learning and audiobook narration where cost-effective, high-quality voiceovers are needed.
  3. 03Accessibility tools, screen readers, and dynamic in-app notifications.
  4. 04Local or edge-device deployment where RAM and compute resources are strictly constrained.
  5. 05Privacy-sensitive applications requiring text-to-speech without exposing proprietary scripts to Big Tech surveillance.

Prompting

Getting better results

Keep text inputs chunked into natural sentences or short paragraphs to maintain consistent pacing and intonation.

Use punctuation (commas, periods, question marks) deliberately to guide the model's natural pauses and inflection.

Select the appropriate language code and voice preset (e.g., 'af_heart') to match your target demographic and tone.

Version history

v0.19
2024-12-25

Initial release with <100 hours of training data and 10 voices.

v1.0
2025-01-27

Major upgrade with 54 voices, 8 languages, and improved ELO.

FAQ

Frequently asked questions

Kokoro is an open-weight, ultra-efficient text-to-speech (TTS) model developed by hexgrad. Despite having only 82 million parameters, it produces highly natural, studio-quality audio that rivals much larger proprietary models.

On Venice, Kokoro is billed at an incredibly low rate of $3.50 per 1 million characters of synthesized speech. This makes it one of the most cost-effective high-quality TTS options available.

Yes, Kokoro is an open-weight model released under the highly permissive Apache-2.0 license. You can freely use, modify, and deploy it for both personal and commercial projects.

While ElevenLabs Turbo v2.5 offers industry-leading expressive speech and instant voice cloning, it is closed-source and costs $62.50 per 1M characters. Kokoro provides comparable naturalness for standard speech at just $3.50 per 1M characters, using open weights.

No, the base Kokoro-82M model does not support native zero-shot voice cloning. Instead, it relies on a curated library of 54 high-quality preset voices spanning multiple accents and languages.

Venice operates under a strict zero-retention policy. When you use Kokoro on Venice, your text prompts are processed anonymously and are never stored, profiled, or used to train future models, maintaining your data sovereignty.

Kokoro v1.0 is a multilingual model supporting 8 languages, including English, Spanish, French, Japanese, and others, with tailored voicepacks for different accents.

Run Kokoro Text to Speech privately

No prompt logging. No data used for training.