Custom Voice Creator 🎛️
Blend 2–5 base Kokoro voices with adjustable weights to create unique new timbres.
Preview instantly, fine-tune the blend, then download the .bin file
for use with a compatible Kokoro TTS engine. Blending and synthesis run in your browser;
non-English Kokoro test text uses the documented phonemization request.
Load Model
Download Kokoro-82M (~305MB) once. Cached in browser.
Pick & Blend
Select 2-5 base voices, adjust weight sliders to taste.
Preview
Type test text and hear the blended voice instantly.
Download
Save the blend as a .bin file for a compatible Kokoro voice loader.
How voice blending works
Each Kokoro voice is stored as a .bin file containing a Float32Array
of style embedding vectors (256 elements per token position). Blending computes a
weighted average: for each position i in the embedding array,
blend[i] = Σ(wⱼ × voiceⱼ[i]) / Σ(wⱼ).
The result is a new .bin file that works immediately with the standard Kokoro-82M ONNX model — no retraining, no fine-tuning. The blend captures characteristics from all source voices proportional to their weights.
Language handling: The dominant base voice (highest weight) determines the language for phonemization. English uses kokoro-js's built-in phonemizer; non-English voices phonemize via the OfflineTTS API. Audio synthesis always runs locally.
Experimental Tool Boundaries
This is an experimental embedding mixer for Kokoro, not a voice-cloning, training, or identity-replication service. It calculates a weighted average of existing compatible style arrays. It does not learn from a microphone sample, infer a real person's vocal identity, or create a new language model. A blend can sound intermediate, unstable, muffled, or unexpectedly similar to one dominant source voice.
Compatible array length and model version matter. A downloaded blend is intended for a matching Kokoro-82M voice-loading workflow; it is not a universal voice file for other TTS engines. Keep the blend recipe and source voice identifiers with the file so you can reproduce or retire it later. The `.bin` output contains style values, not generated speech, a recording, or an automatically granted commercial license.
How to Evaluate a Blend
Begin with two source voices and equal weights. Use a fixed test passage containing short and long sentences, punctuation, names, numbers, questions, and the language you will actually publish. Listen to both source voices first, then the blend at the same model precision, speed, browser, and backend. Change one weight at a time so an improvement can be attributed to a specific adjustment.
Check intelligibility, pronunciation, pacing, loudness changes, high-frequency artifacts, and consistency across several sentences. Do not judge a voice from one greeting. The highest-weight source determines the language path used by this tool, which does not make a cross-language blend natively multilingual. Review non-English Kokoro text handling in the Privacy Policy before testing confidential text.
Output, Rights, and Data Handling
The browser downloads required model and base-voice assets, calculates the blend on the device, previews local speech, and creates the downloadable file without uploading the generated audio. English Kokoro uses its local phonemizer after required assets are available; non-English Kokoro sends entered test text to the documented phonemization endpoint before local synthesis. Model delivery and ordinary site analytics are separate network activity. You are responsible for the upstream terms attached to each source voice, your test text, any person or character the result may evoke, and the context in which generated audio is published.