DictoEaso-Official/offline-tts-models

DictoEaso Offline TTS Models

This repository contains offline TTS assets used by DictoEaso mobile/tablet apps.

It is an asset bundle (models + tokenizers/voices + mapping), not a full TTS runtime. Your app/runtime decides how to load and run each engine (Kokoro / Piper / VITS).

---

What’s inside (Engines & Languages)

1) Kokoro (ONNX FP16) — EN / ES / ZH / JA

• One FP16 ONNX model shared by multiple languages.
• We ship a minimal set of voice binaries (female + male per language) to keep the package small.

Included voices (selected):

• English: af_bella (female), am_adam (male)
• Spanish: ef_dora (female), em_alex (male)
• Chinese: zf_xiaoxiao (female), zm_yunxi (male)
• Japanese: jf_alpha (female), jm_kumo (male)

2) Piper — FR

• Piper-style ONNX model + tokens.txt for French.

3) Korean — VITS / Mimic3 (KO)

• Korean is provided as a VITS/Mimic3 voice package:
  • model.onnx
  • tokens.txt
  • espeak-ng-data/ (phonemizer data required by many Mimic3/VITS pipelines)

Note: KO is not the same engine as Piper or Kokoro. Treat it as a separate runtime path.

---

Repository layout

. ├─ kokoro/ │ └─ core_v1_fp16/ │ ├─ onnx/ │ │ └─ model_fp16.onnx │ ├─ voices/ │ │ ├─ af_bella.bin │ │ ├─ am_adam.bin │ │ ├─ ef_dora.bin │ │ ├─ em_alex.bin │ │ ├─ zf_xiaoxiao.bin │ │ ├─ zm_yunxi.bin │ │ ├─ jf_alpha.bin │ │ └─ jm_kumo.bin │ ├─ config.json │ ├─ tokenizer.json │ └─ tokenizer_config.json │ ├─ piper/ │ └─ fr/ │ ├─ model.onnx │ └─ tokens.txt │ ├─ vits/ │ └─ ko_kss_low/ │ ├─ model.onnx │ ├─ tokens.txt │ └─ espeak-ng-data/ │ └─ (many files) │ └─ voice_map.json

pgsql Copy code

---

voice_map.json

voice_map.json provides a stable mapping from language → engine → paths + default voices.

Example shape (your app should follow the real file in this repo):

{
  "kokoro": {
    "core": "kokoro/core_v1_fp16/onnx/model_fp16.onnx",
    "voicesDir": "kokoro/core_v1_fp16/voices",
    "defaults": {
      "en": {"female":"af_bella","male":"am_adam"},
      "es": {"female":"ef_dora","male":"em_alex"},
      "zh": {"female":"zf_xiaoxiao","male":"zm_yunxi"},
      "ja": {"female":"jf_alpha","male":"jm_kumo"}
    }
  },
  "piper": {
    "fr": {"model":"piper/fr/model.onnx","tokens":"piper/fr/tokens.txt"}
  },
  "vits": {
    "ko": {
      "model":"vits/ko_kss_low/model.onnx",
      "tokens":"vits/ko_kss_low/tokens.txt",
      "espeakData":"vits/ko_kss_low/espeak-ng-data"
    }
  }
}
How to download (recommended)
Using Hugging Face CLI (hf):

bash
Copy code
hf download DictoEaso-Official/offline-tts-models --local-dir ./offline-tts-models --local-dir-use-symlinks False
If you only want Kokoro core:

bash
Copy code
hf download DictoEaso-Official/offline-tts-models \
  --include "kokoro/**" "voice_map.json" \
  --local-dir ./offline-tts-models --local-dir-use-symlinks False
Notes for app/runtime integration
Kokoro
Load ONNX: kokoro/core_v1_fp16/onnx/model_fp16.onnx

Load tokenizer + config from the same folder

Pick voice by .bin name (without extension), e.g. af_bella, jm_kumo

Piper (FR)
Load model: piper/fr/model.onnx

Load tokens: piper/fr/tokens.txt

VITS/Mimic3 (KO)
Load model: vits/ko_kss_low/model.onnx

Load tokens: vits/ko_kss_low/tokens.txt

Many pipelines require espeak-ng-data/ to phonemize → ensure your KO runtime knows how to use it.

Licenses / Third-party notices (IMPORTANT)
This repository aggregates multiple upstream projects and does not have a single unified license for all files.

The repo-level metadata uses license: other to reflect that the contents are mixed.

Each subcomponent has its own upstream license and attribution requirements.

Upstream sources (you must verify before redistribution)
Kokoro ONNX: derived from upstream onnx-community/Kokoro-82M-v1.0-ONNX (model + tokenizer/config + voices).

Piper FR: Piper-compatible voice (ONNX + tokens). License depends on the specific voice origin.

VITS/Mimic3 KO (kss_low): Mimic3/VITS voice package. License and attribution requirements are defined by its upstream distribution (and may differ from Kokoro/Piper).

Internal use tip: If your legal/compliance requires strict separation, consider splitting KO into a dedicated repo so licensing boundaries remain crystal-clear.

Change log policy
kokoro/core_v1_fp16/*: updated only when upstream Kokoro ONNX or selected voices change

piper/fr/*: updated per voice refresh

vits/ko_kss_low/*: updated per upstream KO voice refresh

voice_map.json: kept stable; changes are backward compatible when possible