Kokoro TTS

Run Kokoro TTS — an ultra-lightweight 82M-parameter text-to-speech model on Clore.ai GPUs.

Kokoro is an 82M-parameter text-to-speech model that punches far above its weight class. Despite its tiny size (under 2 GB VRAM), it produces remarkably natural English speech and runs at real-time or faster speeds even on budget hardware. With Apache 2.0 licensing, multiple built-in voice styles, and CPU inference support, Kokoro is ideal for real-time applications, chatbots, and edge deployments.

HuggingFace: hexgrad/Kokoro-82M PyPI: kokoro License: Apache 2.0

Key Features

• 82M parameters — one of the smallest high-quality TTS models available

• < 2 GB VRAM — runs on virtually any GPU, and even on CPU

• Multiple voice styles — American English, British English; male and female voices

• Real-time or faster — low-latency inference suitable for streaming

• Streaming generation — yields audio chunks as they are produced

• Multi-language support — English (primary), Japanese (misaki[ja]), Chinese (misaki[zh])

• Apache 2.0 — free for personal and commercial use

Requirements

Component	Minimum	Recommended
GPU	Any with 2 GB VRAM	RTX 3060
VRAM	2 GB	4 GB
RAM	4 GB	8 GB
Disk	500 MB	1 GB
Python	3.9+	3.11
System	espeak-ng installed	—

Clore.ai recommendation: An RTX 3060 ($0.03–0.07/hr) is more than enough. Kokoro can even run on CPU-only instances for extremely cost-effective TTS.

Installation

# Install system dependency
apt-get install -y espeak-ng

# Install Kokoro and audio I/O
pip install kokoro>=0.9.4 soundfile torch

# For Japanese support (optional)
pip install misaki[ja]

# For Chinese support (optional)
pip install misaki[zh]

# Verify
python -c "from kokoro import KPipeline; print('Kokoro ready')"

Quick Start

from kokoro import KPipeline
import soundfile as sf

# Initialize pipeline
# 'a' = American English, 'b' = British English
pipeline = KPipeline(lang_code='a')

text = """
Kokoro is a lightweight text-to-speech model with only eighty-two million
parameters. Despite its small size, it produces natural and expressive speech.
"""

# Generate audio — voice options: af_heart, af_bella, af_nicole, af_sarah, af_sky,
#                                  am_adam, am_michael, bf_emma, bf_isabella, bm_george, bm_lewis
generator = pipeline(text, voice='af_heart', speed=1.0)

for i, (graphemes, phonemes, audio) in enumerate(generator):
    sf.write(f'output_{i}.wav', audio, 24000)
    print(f"Chunk {i}: {graphemes[:50]}...")

print("Done!")

Usage Examples

Multiple Voices Comparison

Generate the same text with different voices to compare:

from kokoro import KPipeline
import soundfile as sf

pipeline = KPipeline(lang_code='a')

text = "Welcome to Clore.ai, the peer-to-peer GPU marketplace."

voices = ['af_heart', 'af_bella', 'am_adam', 'am_michael']

for voice in voices:
    generator = pipeline(text, voice=voice, speed=1.0)
    for i, (gs, ps, audio) in enumerate(generator):
        sf.write(f'{voice}_{i}.wav', audio, 24000)
    print(f"Generated: {voice}")

British English with Speed Control

from kokoro import KPipeline
import soundfile as sf

# 'b' = British English
pipeline = KPipeline(lang_code='b')

text = "Good afternoon. This is a demonstration of British English synthesis."

# speed < 1.0 = slower, speed > 1.0 = faster
generator = pipeline(text, voice='bf_emma', speed=0.85)

all_audio = []
for gs, ps, audio in generator:
    all_audio.append(audio)

import numpy as np
combined = np.concatenate(all_audio)
sf.write('british_slow.wav', combined, 24000)
print(f"Total duration: {len(combined)/24000:.1f}s")

Batch File Processing

Process multiple texts and concatenate into a single audiobook-style file:

from kokoro import KPipeline
import soundfile as sf
import numpy as np

pipeline = KPipeline(lang_code='a')

chapters = [
    "Chapter one. The beginning of our journey starts here.",
    "The sun rose over the mountains, casting long shadows across the valley.",
    "She opened the door and stepped into the unknown.",
]

all_audio = []
silence = np.zeros(int(24000 * 0.5))  # 0.5s silence between chapters

for idx, text in enumerate(chapters):
    for gs, ps, audio in pipeline(text, voice='af_bella', speed=1.0):
        all_audio.append(audio)
    all_audio.append(silence)
    print(f"Chapter {idx+1} done")

combined = np.concatenate(all_audio)
sf.write('audiobook.wav', combined, 24000)
print(f"Total: {len(combined)/24000:.1f}s")

Tips for Clore.ai Users

• CPU inference — Kokoro is small enough to run on CPU; useful for cost-sensitive workloads or when GPUs are unavailable

• Streaming — the generator yields audio chunks as they are produced, enabling real-time playback in web apps

• Combine with WhisperX — use WhisperX for transcription and Kokoro for re-synthesis in voice pipelines

• Docker — use pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime and add apt-get install -y espeak-ng to your startup

• Voice consistency — stick to one voice ID per project for a consistent narrator experience

• Cost efficiency — at $0.03–0.07/hr on an RTX 3060, Kokoro is one of the cheapest TTS solutions to self-host

Troubleshooting

Problem	Solution
espeak-ng not found	Run apt-get install -y espeak-ng (required system dependency)
ModuleNotFoundError: kokoro	Install with pip install kokoro>=0.9.4 soundfile
Audio sounds robotic	Try a different voice (e.g., af_heart tends to sound most natural)
Japanese/Chinese not working	Install language extras: pip install misaki[ja] or misaki[zh]
Out of memory on CPU	Reduce text length per call; Kokoro streams chunks so memory stays bounded
Slow first run	Model weights download on first use (~200 MB); subsequent runs are instant