Logo

KaniTTS

A high-speed, high-fidelity Text-to-Speech model optimized for real-time conversational AI applications.

Overview

KaniTTS uses a two-stage pipeline combining a large language model with an efficient audio codec for exceptional speed and audio quality. The architecture generates compressed token representations through a backbone LLM, then rapidly synthesizes waveforms via neural audio codec, achieving extremely low latency.

Key Specifications:

Model Size: 370M parameters
Sample Rate: 22kHz
Languages: English, German, Chinese, Korean, Arabic, Spanish
License: Apache 2.0

Performance

Nvidia RTX 5080 Benchmarks:

Latency: ~1 second to generate 15 seconds of audio
Memory: 2GB GPU VRAM
Quality Metrics: MOS 4.3/5 (naturalness), WER <5% (accuracy)

Pretraining:

Dataset: ~80k hours from LibriTTS, Common Voice, and Emilia
Hardware: 8x H100 GPUs, 45 hours training time on Lambda AI

Voices Datasets

Voices:

david — David, English (British)
puck — Puck, English (Gemini)
kore — Kore, English (Gemini)
andrew — Andrew, English
jenny — Jenny, English (Irish)
simon — Simon, English
katie — Katie, English
seulgi — Seulgi, Korean
bert — Bert, German
thorsten — Thorsten, German (Hessisch)
maria — Maria, Spanish
mei — Mei, Chinese (Cantonese)
ming — Ming, Chinese (Shanghai OpenAI)
karim — Karim, Arabic
nur — Nur, Arabic

Quickstart: Install from PyPI & Run Inference

It’s a lightweight so you can install, load a model, and speak in minutes. Designed for quick starts and simple workflows—no heavy setup, just pip install and run. More detailes...

Install

pip install kani-tts
pip install -U "transformers==4.57.1" # for LFM2 !!!

Quick Start

from kani_tts import KaniTTS

model = KaniTTS('nineninesix/kani-tts-370m')

# Generate audio from text
audio, text = model("Hello, world!")

# Save to file (requires soundfile)
model.save_audio(audio, "output.wav")

Working with Multi-Speaker Models

This model support multiple speakers. You can check if your model supports speakers and select a specific voice:

from kani_tts import KaniTTS

model = KaniTTS('nineninesix/kani-tts-370m')

# Check if model supports multiple speakers
print(f"Model type: {model.status}")  # 'singlspeaker' or 'multispeaker'

# Display available speakers (pretty formatted)
model.show_speakers()

# Or access the speaker list directly
print(model.speaker_list)  # ['andrew', 'katie', ...]

# Generate audio with a specific speaker
audio, text = model("Hello, world!", speaker_id="andrew")

Custom Configuration

from kani_tts import KaniTTS

model = KaniTTS(
    'nineninesix/kani-tts-370m',
    temperature=0.7,           # Control randomness (default: 1.0)
    top_p=0.9,                 # Nucleus sampling (default: 0.95)
    max_new_tokens=2000,       # Max audio length (default: 1200)
    repetition_penalty=1.2,    # Prevent repetition (default: 1.1)
    suppress_logs=True,        # Suppress library logs (default: True)
    show_info=True,            # Show model info on init (default: True)
)

audio, text = model("Your text here")

Playing Audio in Jupyter Notebooks

You can listen to generated audio directly in Jupyter notebooks or IPython:

from kani_tts import KaniTTS
from IPython.display import Audio as aplay

model = KaniTTS('nineninesix/kani-tts-370m')
audio, text = model("Hello, world!")

# Play audio in notebook
aplay(audio, rate=model.sample_rate)

Audio Examples

Text	Audio
I do believe Marsellus Wallace, MY husband, YOUR boss, told you to take me out and do WHATEVER I WANTED.
What do we say to the god of death? Not today!
What do you call a lawyer with an IQ of 60? Your honor
You mean, let me understand this cause, you know maybe it's me, it's a little fucked up maybe, but I'm funny how, I mean funny like I'm a clown, I amuse you?

Use Cases

Conversational AI: Real-time speech for chatbots and virtual assistants
Edge/Server Deployment: Resource-efficient inference on affordable hardware
Accessibility: Screen readers and language learning applications
Research: Fine-tuning for specific voices, accents, or emotions

Limitations

Performance degrades with inputs exceeding 2000 tokens
Limited expressivity without fine-tuning for specific emotions
May inherit biases from training data in prosody or pronunciation
Optimized primarily for English; other languages may require additional training

Optimization Tips

Multilingual Performance: Continually pretrain on target language datasets and fine-tune NanoCodec
Batch Processing: Use batches of 8-16 for high-throughput scenarios
Hardware: Optimized for NVIDIA Blackwell architecture GPUs

Resources

Models:

Examples:

Links:

Acknowledgments

Built on top of LiquidAI LFM2 350M as the backbone and Nvidia NanoCodec for audio processing.

Responsible Use

Prohibited activities include:

Illegal content or harmful, threatening, defamatory, or obscene material
Hate speech, harassment, or incitement of violence
Generating false or misleading information
Impersonating individuals without consent
Malicious activities such as spamming, phishing, or fraud

By using this model, you agree to comply with these restrictions and all applicable laws.

Contact

Have a question, feedback, or need support? Please fill out our contact form and we'll get back to you as soon as possible.

Citation

@misc {sb_2025,
    author       = { SB },
    title        = { gemini-flash-2.0-speech },
    year         = 2025,
    url          = { https://huggingface.co/datasets/shb777/gemini-flash-2.0-speech },
    doi          = { 10.57967/hf/4237 },
    publisher    = { Hugging Face }
}

@misc{toyin2025arvoicemultispeakerdatasetarabic,
      title={ArVoice: A Multi-Speaker Dataset for Arabic Speech Synthesis}, 
      author={Hawau Olamide Toyin and Rufael Marew and Humaid Alblooshi and Samar M. Magdy and Hanan Aldarmaki},
      year={2025},
      eprint={2505.20506},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.20506}, 
}

@misc {thorsten_müller_2024,
    author       = { {Thorsten Müller} },
    title        = { TV-44kHz-Full (Revision ff427ec) },
    year         = 2024,
    url          = { https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full },
    doi          = { 10.57967/hf/3290 },
    publisher    = { Hugging Face }
}

@misc{carlosmenaciempiessfem2019,
      title={CIEMPIESS FEM CORPUS: Audio and Transcripts of Female Speakers in Spanish.}, 
      ldc_catalog_no={LDC2019S07},
      DOI={https://doi.org/10.35111/xdx5-n815},
      author={Hernandez Mena, Carlos Daniel},
      journal={Linguistic Data Consortium, Philadelphia},
      year={2019},
      url={https://catalog.ldc.upenn.edu/LDC2019S07},
}

Downloads last month: 4,347

Safetensors

Model size

0.4B params

Tensor type

F32

Model tree for nineninesix/kani-tts-370m

Base model

nineninesix/kani-tts-450m-0.2-pt

Finetuned

(4)

this model

Finetunes

2 models

Quantizations

3 models

nineninesix
/

kani-tts-370m