Custom Voice Tester 🎛️

66 custom blended voices created by mixing existing Kokoro TTS voice embeddings. Each blend combines 2–5 base voices with different weights to create unique timbres. 20+ Mandarin Chinese voices are included, plus English, Japanese, European, and fusion blends.

①

Load Model

Download Kokoro-82M (~305MB) once. Cached in browser.

②

Type Text

Enter any text to test. English is 100% offline; other languages phonemize via API.

③

Test Voices

Click Play on any voice card. Or Play All to hear every voice sequentially.

How custom voice blending works

Each Kokoro TTS voice is stored as a .bin file containing a Float32Array of style embeddings at 256-element intervals. When blending, we take a weighted average of the embedding vectors from multiple source voices, producing a brand-new voice that combines characteristics of all source voices proportionally.

The blend script (scripts/generate_custom_voices.py) loads source voice .bin files, computes the weighted average, and saves the result as a new .bin file. No retraining or fine-tuning is needed — the result works immediately with the standard Kokoro-82M ONNX model.

Language handling: The test page detects the dominant source voice's language and uses it for phonemization. For English blends, kokoro-js's built-in espeak phonemizer is used. For non-English blends (Chinese, Japanese, etc.), text is phonemized via the OfflineTTS phonemize API, which runs misaki (Japanese/Chinese) or espeak-ng (other languages). Audio synthesis always happens locally on your device.

API: POST /api/phonemize with { text, lang } → returns IPA phonemes. English codes: a, b. Non-English: j (ja), z (zh), e (es), f (fr), h (hi), i (it), p (pt-br).