Miko voice for piper-plus

A small Japanese text-to-speech model that speaks as Miko, the official character of AITuber OnAir. It is a piper-plus VITS model fine-tuned from the multilingual zs-v1 base, exported to ONNX in FP16 so it runs in the browser with WebAssembly.

The voice is synthetic. The training audio was generated with Irodori-TTS from a designed voice; no person was recorded.

Two versions are available. sample-voice-v2 is the recommended one; it is closer to the teacher voice in intonation and tempo and makes fewer reading errors, at the same size and with the same inputs as v1.

Files

• Sample rate: 22,050 Hz, mono
• Inputs: input, input_lengths, scales, lid, prosody_features
• Language id: ja = 0 (the model is based on a six-language table but was fine-tuned on Japanese only)
• Suggested length_scale: 1.07 for v2, 1.1 for v1 (slightly slower than each model's native pace)

Usage

With @aituber-onair/voice 0.27.0 or later, which builds the padded phoneme sequence this model expects:

import { PiperPlusEngine } from '@aituber-onair/voice';

const engine = new PiperPlusEngine();
engine.setAssets({
  basePath: '/piper/', // piper-plus web runtime + OpenJTalk dictionary
  modelConfigFile: 'sample-voice-v2.onnx.json',
  modelFile: 'sample-voice-v2.onnx',
  voiceFile: 'mei_normal.htsvoice',
});
engine.setSpeed(0.93); // length_scale = 1 / speed ≈ 1.07
engine.setNoiseScale(0.667);
const wav = await engine.fetchAudio({ style: 'talk', message: 'こんにちは' }, '');

The piper-plus web runtime (npm install piper-plus) can load the same files through its multilingual code path.

Training

Both versions:

• Base: piper-plus zs-v1 (ayousanz/piper-plus-base, releases/zs-v1)
• Fine-tuned as a single speaker on synthetic Japanese speech; training sentences are original sentences plus the public-domain ITA corpus
• Exported with piper-plus 2.0.0

v2:

• About 3.8 hours of training audio (3,674 sentences), about 4.6 times v1
• Training phoneme sequences rebuilt in the same padded format the inference path produces (v1 was trained on an unpadded format, which made it slower and pulled its intonation toward the base model)
• Speaker conditioning folded into the language embedding before fine-tuning, and the duration predictor trained instead of frozen
• 120 epochs

v1:

• About 0.8 hours of training audio (734 sentences)
• 300 epochs, duration predictor frozen

Evaluation

Averages over 50 held-out test sentences, compared with the teacher audio.

v2 still sounds slightly rougher than the teacher.

License and notices

• Model weights: MIT, following the base model. See BASE_MODEL_LICENSE.txt and BASE_MODEL_NOTICE.md for the base model's provenance, including the datasets its authors used.
• Miko is the character of AITuber OnAir, © Yuki Shindo. Use of the character follows the Miko Character Usage Guidelines.
• Do not present output from this model as a recording of a real person.