README: model journey section with diagram
Browse files
README.md
CHANGED
|
@@ -68,6 +68,12 @@ audio = model(phonemes, ref_s, speed=1.0) # 24 kHz mono tensor
|
|
| 68 |
|
| 69 |
`kokoro_generate.py` implements all three with auto-detection.
|
| 70 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
## Training
|
| 72 |
|
| 73 |
Two-stage StyleTTS2 finetune (via the [kikiri-tts](https://github.com/semidark/kikiri-tts) recipe) from Kokoro-82M v1.0 weights, multi-speaker (15 speakers):
|
|
|
|
| 68 |
|
| 69 |
`kokoro_generate.py` implements all three with auto-detection.
|
| 70 |
|
| 71 |
+
## Model journey
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
Three iterations led here: **Qwen3-TTS 1.7B** (good pronunciation, noisy output, one checkpoint per persona — dropped) → **VoxCPM2 ~800M LoRA** (excellent cloning quality, but needs a reference clip at inference and server-class compute) → **Kokoro-82M** (this model: tiny, edge/CPU-capable, personas as 512KB voicepacks, zero reference audio).
|
| 76 |
+
|
| 77 |
## Training
|
| 78 |
|
| 79 |
Two-stage StyleTTS2 finetune (via the [kikiri-tts](https://github.com/semidark/kikiri-tts) recipe) from Kokoro-82M v1.0 weights, multi-speaker (15 speakers):
|