humair025/kikiri-urdu-stage1_2run

Kikiri-tts Urdu finetuning (first_styletts2 stage)

This repo was auto-uploaded by pipeline.checkpointer.KikiriCheckpointer.

• Stage: first_styletts2
• Final step: 6320
• Final epoch: 1
• Best val loss: 0.6203 (at step 6000)
• Pipeline version: 0.1.0

Layout:

• checkpoints/final.pth — final StyleTTS2-format checkpoint (loadable via scripts/test_inference.py --checkpoint).
• checkpoints/best.pth — checkpoint with lowest val loss.
• checkpoints/step_*.pth — periodic checkpoints.
• tensorboard_logs/ — TensorBoard event files.
• config.yaml — training config snapshot.

Inference

# After installing the kikiri-tts repo + StyleTTS2 submodule:
from kokoro import KModel, KPipeline
import torch, soundfile

model = KModel(config="training/config.json",
               model="checkpoints/final.pth")
pipeline = KPipeline(lang_code="d", model=model)  # 'd' for German config;
                                                   # adjust for Urdu
voice = torch.load("voices/<voice>.pt", map_location="cpu")
for text in ["یار، بس پھر!"]:
    for gs, ps, audio in pipeline(text, voice=voice, speed=1):
        soundfile.write("out.wav", audio, 24000)