hexgrad/Kokoro-82M · Hugging Face

Overview

Kokoro-82M is a text-to-speech (TTS) model developed by Hexgrad, available on Hugging Face under the model ID hexgrad/Kokoro-82M. With 82 million parameters, it is a relatively lightweight yet powerful TTS model. It has achieved significant popularity, amassing over 16.7 million downloads. The model is based on the StyleTTS2 architecture, specifically fine-tuned from yl4579/StyleTTS2-LJSpeech. It supports English and Arabic languages and is released under the Apache 2.0 license.

According to the official model card, Kokoro is an open-weight TTS model that delivers comparable quality to larger models while being significantly faster and more cost-efficient. The model's weights are Apache-licensed, allowing deployment anywhere from production environments to personal projects. The official GitHub repository is at https://github.com/hexgrad/kokoro, and a live demo is available at https://hf.co/spaces/hexgrad/Kokoro-TTS.

As of April 2025, the market rate for Kokoro served over API is under $1 per million characters of text input, or under $0.06 per hour of audio output. On average, 1000 characters of input corresponds to about 1 minute of output. The model has been deployed in numerous projects and commercial APIs.

Capabilities

Kokoro-82M excels in generating natural-sounding speech from text. Key capabilities include:

• Multilingual Support: Primarily English and Arabic. The model card lists training data for v1.0 as "Few hundred hrs" with 8 languages and 54 voices, while v0.19 had less than 100 hours with 1 language and 10 voices.
• High-Quality Speech: Leverages the StyleTTS2 framework for expressive and realistic voice synthesis.
• Efficient Inference: With 82M parameters, it balances quality and computational efficiency, suitable for real-time applications.
• Customizable Voice Styles: Inherits StyleTTS2's ability to control speaking style, emotion, and prosody.
• Open Source: Apache 2.0 license allows for commercial and research use.
• Cost-Effective: The model is designed to be significantly faster and more cost-efficient than larger models, with API rates under $1 per million characters.

Use cases

• Voice Assistants: Integrate into virtual assistants for natural interaction.
• Accessibility Tools: Convert text to speech for visually impaired users.
• Content Creation: Generate voiceovers for videos, podcasts, and audiobooks.
• Language Learning: Provide pronunciation examples for English and Arabic learners.
• Customer Service: Automate voice responses in call centers.
• Production Deployments: The Apache license allows commercial use, and the model has been deployed in numerous projects and commercial APIs.

License & deployment

• License: Apache 2.0, permitting free use, modification, and distribution. The model card explicitly states: "This is an Apache-licensed model, and Kokoro has been deployed in numerous projects and commercial APIs. We welcome the deployment of the model in real use cases."
• Deployment: Can be deployed via Hugging Face Inference API, or locally using libraries like transformers and torch. The model is optimized for GPU inference but can run on CPU with reduced speed. The official usage example uses kokoro Python package and espeak-ng for phonemization.
• Hardware Considerations: Training costs are reported as 1000 A100 80GB GPU hours total (500 for v0.19 and 500 for v1.0), with an average hourly rate of $1/hour, totaling $1000. Inference is efficient on modern GPUs.
• Safety Notes: The model card warns about fake websites like kokorottsai.com and kokorotts.net that are not affiliated with the model. Users should only use official sources.

Alternatives

• StyleTTS2: The base model, also open-source, with similar capabilities.
• Tacotron2: A classic TTS model, widely used but less expressive.
• FastSpeech2: Non-autoregressive model, faster inference.
• VITS: End-to-end TTS with variational inference, high quality.
• Coqui TTS: Open-source TTS toolkit with multiple models.
• Other Kokoro Variants: The model card mentions v0.19 and v1.0 releases, with v1.0 supporting more languages and voices.

FAQ

Q: What languages does Kokoro-82M support? A: It supports English and Arabic according to the overview, but the model card for v1.0 mentions 8 languages and 54 voices. The exact language list is not fully specified in the source; users should refer to the official model card for details.

Q: Is Kokoro-82M free to use? A: Yes, it is released under Apache 2.0 license, free for commercial and non-commercial use.

Q: How can I use Kokoro-82M? A: You can use it via Hugging Face's transformers library, the kokoro Python package, or directly from the Hugging Face Hub. The official usage example requires kokoro>=0.9.2, soundfile, and espeak-ng.

Q: What is the model architecture? A: It is based on StyleTTS2, fine-tuned from yl4579/StyleTTS2-LJSpeech.

Q: What are the training costs? A: Training used 1000 A100 80GB GPU hours, with an estimated cost of $1000.

Q: Are there any known scams? A: Yes, the model card warns about fake websites like kokorottsai.com and kokorotts.net that are not affiliated with the model. Always use official sources.

Q: Can I deploy Kokoro-82M commercially? A: Yes, the Apache 2.0 license permits commercial use.

Q: What is the audio quality? A: The model delivers quality comparable to larger models, with 24kHz sample rate in the demo.

Q: How do I cite Kokoro-82M? A: The model card provides a DOI: 10.57967/hf/4329, and references arXiv papers 2306.07691 and 2203.02395.