AkinVox Kokoro Cloning v1

English voice cloning for Kokoro. Give it a short recording, its transcript and new text. Turn the adapter off to use the original Kokoro voices.

I’m building this for my audiobook site, akinvox.com. The goal is voice cloning without sacrificing quality or speed. V1 has had 100k cloning updates on pretrained Kokoro; I believe more training could improve it substantially. Work on v2 is ongoing. The site is still under construction, and I’d appreciate you giving it a go.

GitHub & guide · Weights

Samples

Each example uses a different synthetic voice and passage. Compare the reference, its same-sentence clone, then a clone speaking different words.

Strengths

Five voices across narration, dialogue, a longer passage, explanation and questions. These are selected examples; voice similarity varies with the reference.

Warm narration · Warm British woman

Listen for the voice character and relaxed storytelling rhythm.

1. Reference

Transcript

The morning sunlight filled the room. I opened the window, took a quiet breath, and smiled at the beautiful day ahead.

2. Clone · same sentence

Transcript

The morning sunlight filled the room. I opened the window, took a quiet breath, and smiled at the beautiful day ahead.

3. Clone · different sentence

Transcript

The old bookshop stood at the end of a narrow street. Inside, every shelf held the promise of a different adventure.

Lively dialogue · Bright American woman

Listen for questions, replies and changing emphasis. One voice reads both characters.

1. Reference

Transcript

Did you leave the tickets on the table? she asked. I put them in your bag, he replied. Then we had better catch that train.

2. Clone · same sentence

Transcript

Did you leave the tickets on the table? she asked. I put them in your bag, he replied. Then we had better catch that train.

3. Clone · different sentence

Transcript

You brought the map? she asked. Of course I did, he replied. Then why are we walking in circles? For a moment, neither of them said a word.

A longer passage · Low, mature British man

Listen for a consistent voice and clear phrasing through a sustained passage.

1. Reference

Transcript

The ferry left the harbour just after dawn. Daniel watched the houses shrink behind him. By the time the sun reached the hills, the island was only a distant line.

2. Clone · same sentence

Transcript

The ferry left the harbour just after dawn. Daniel watched the houses shrink behind him. By the time the sun reached the hills, the island was only a distant line.

3. Clone · different sentence

Transcript

Rain tapped softly against the window. Daniel set his cup beside the open book and turned another page. Outside, the town was waking up. A delivery van stopped near the bakery, and someone called a greeting across the street. He had planned to leave an hour ago, but the letter on the table had changed everything. He read the final line once more, folded the paper carefully, and reached for his coat. This time, he knew exactly where he needed to go.

Factual explanation · Clear American man

Listen for distinct words, numbers and steady instructional delivery.

1. Reference

Transcript

A compass needle points towards magnetic north. When you turn the map to match it, the landmarks around you become much easier to identify.

2. Clone · same sentence

Transcript

A compass needle points towards magnetic north. When you turn the map to match it, the landmarks around you become much easier to identify.

3. Clone · different sentence

Transcript

Set the timer for twenty minutes and place the tray on the middle shelf. The warm air circulates around the food, helping it cook evenly. Leave a little space between each piece so the heat can reach every side.

Questions and answers · Light, youthful American man

Listen for changes in intonation between questions and answers.

1. Reference

Transcript

Why does the moon look larger near the horizon? Is it really closer to us? The answer has more to do with perception than distance.

2. Clone · same sentence

Transcript

Why does the moon look larger near the horizon? Is it really closer to us? The answer has more to do with perception than distance.

3. Clone · different sentence

Transcript

Is the bridge still open? Can we cross before sunset? Yes, said the guide, but we should leave now. What happens if the weather changes? Then we take the sheltered path through the trees.

Quick start

Quick-start example · Stock Kokoro af_heart

The synthetic reference included in the installation guide.

1. Reference

Transcript

I love getting lost in a good audiobook. Every story opens a new world, and a familiar voice makes the journey feel like home.

2. Clone · same sentence

Transcript

I love getting lost in a good audiobook. Every story opens a new world, and a familiar voice makes the journey feel like home.

3. Clone · different sentence

Transcript

Welcome to AkinVox. Today, we are trying a new voice and telling a different story.

Weaknesses

Expressive delivery, whispering and accents remain difficult.

Expressive delivery · Deep, quiet American man

Pace, breathiness and emphasis can drift.

1. Reference

Transcript

Rain is falling outside the old library. The shelves are full of stories, and a small lamp lights the page. Another chapter is waiting to begin.

2. Clone · same sentence

Transcript

Rain is falling outside the old library. The shelves are full of stories, and a small lamp lights the page. Another chapter is waiting to begin.

3. Clone · different sentence

Transcript

The museum opens early on Saturday. Visitors can explore the new exhibition and discover how these remarkable inventions changed everyday life.

Whisper preservation · Whispering woman

Whispered delivery can become normally voiced speech.

1. Reference

Transcript

A thin layer of snow covers the garden. Beyond the glass, a robin rests on the fence. The whole morning feels still and quiet.

2. Clone · same sentence

Transcript

A thin layer of snow covers the garden. Beyond the glass, a robin rests on the fence. The whole morning feels still and quiet.

3. Clone · different sentence

Transcript

The clock in the hallway has stopped. I can hear the soft rustle of pages as someone turns to the next chapter. Outside, the last light is fading.

Accent preservation · Australian man

The original accent may not carry over.

1. Reference

Transcript

I parked the car near the water, then walked down to the beach. We can meet after work and take the boat home.

2. Clone · same sentence

Transcript

I parked the car near the water, then walked down to the beach. We can meet after work and take the boat home.

3. Clone · different sentence

Transcript

After work, we drove past the harbour and stopped for a walk along the water. The boat was already gone, so we bought a coffee and watched the clouds roll in.

Usage

Python 3.10–3.12, Linux or WSL. CPU works; GPU setup and the local browser demo are in the guide.

Install

git clone https://github.com/akinvox/kokoro-cloning.git
cd kokoro-cloning
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install torch==2.6.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cpu
python -m pip install -c constraints-tested.txt .
python -m spacy download en_core_web_sm

For best results, use 3–10 seconds of one speaker, with at least 3 seconds of actual speech and a 10-second maximum for the final clip. Avoid music, other speakers, clipping and heavy reverb.

The reference-preparation workflow used in AkinVox production is DeepFilterNet3 denoising at 48 kHz → 24 kHz mono → Silero VAD speech trimming → final normalization. Pauses longer than one second are shortened to about 300 ms; normalization targets RMS 0.1 with a 0.99 peak ceiling. Select a complete speech or quiet boundary within 10 seconds.

Transcribe the final cleaned/cropped file, not the original longer recording. Put those exact words, including audible “um” and “uh,” in reference.txt. CrisperWhisper can draft the transcript; check it by listening. Prepare the file before calling the package: denoising/VAD/ASR are not automatically bundled into the public CLI. Full reference-preparation guide.

kokoro-clone --reference reference.wav --reference-text-file reference.txt --text "A new story begins here." --output clone.wav

Original Kokoro, adapter off:

kokoro-clone --adapter off --voice af_heart --text "A new story begins here." --output stock.wav

Release notes

• English only, 24 kHz mono. Long passages work best split into sentences.
• LoRA adapter, reference mapper and reference encoders included. The loader downloads the original base; no merging needed.
• Training code will follow when the research is ready.

Apache-2.0 for AkinVox code and weights, including commercial use. Upstream notices · Validation · Sample details