kikiri-turkish-base

Kikiri Turkish Base

Turkish Kokoro-82M / StyleTTS2 model trained from hexgrad/Kokoro-82M.

using https://github.com/semidark/kikiri-tts

Training data

Dataset: afkfatih/turkish-tts-combined-raw

Prepared corpus:

• 28,958 accepted clips
• 53.021 hours
• 24 kHz mono audio
• Three speakers
• Turkish text normalized and phonemized with eSpeak-ng
• Unsupported IPA ɫ normalized to Kokoro-compatible l

Speakers:

• tf_nisan_kumru: 8,036 samples
• tm_ahmet_deniz: 11,288 samples
• tm_mazlum_kiper: 9,634 samples

Stage 1

Stage 1 was trained on all three speakers for 10 epochs.

Final validation mel loss: approximately 0.267.

checkpoints/stage1/first_stage.pth is the raw StyleTTS2 training checkpoint.

Stage 2

Stage 2 specializes the model for the Nisan Kumru voice.

The recommended checkpoint is the preserved original Epoch 4 model:

• Validation loss: approximately 0.3334
• WavLM adversarial training active
• Natural Turkish pronunciation
• Foreign-accent behavior from Stage 1 substantially reduced

Files:

• checkpoints/stage2/epoch_2nd_00006.pth: raw StyleTTS2 Stage 2 training checkpoint
• models/nisan_epoch7_original.pth: Kokoro-compatible inference model
• voices/nisan_stage2_epoch7_original.pt: Nisan voicepack with acoustic and prosodic style vectors

The word stripped means that the DataParallel module. prefix was removed from checkpoint state-dictionary keys. Model parameters were not pruned.

Audio samples

The following samples were generated using the Epoch 7 Nisan model and nisan_stage2_epoch7.pt voicepack at 24 kHz with speed=1.0.

Merhaba

Text: Merhaba, bugün hava gerçekten çok güzel.

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_01.wav" type="audio/wav"></audio>

İstanbul–Ankara

Text: İstanbul'dan Ankara'ya hızlı trenle seyahat ediyoruz.

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_02.wav" type="audio/wav"></audio>

Model quality

Text: Bu model Türkçe konuşmayı artık daha iyi öğreniyor.

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_03.wav" type="audio/wav"></audio>

Turkish phonetic coverage

Text: Şu küçük çocuk, öğleden sonra üzüm ve çilek yiyecek.

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_04.wav" type="audio/wav"></audio>

Question intonation

Text: Yarın bizimle birlikte sinemaya gelmek ister misin?

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_05.wav" type="audio/wav"></audio>

Text: Uzun

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_06.wav" type="audio/wav"></audio>

<audio controls><source src="https://huggingface.co/duxx/kikiri-turkish-base/resolve/main/samples/epoch7/test_07.wav" type="audio/wav"></audio>

Base model

This model is derived from hexgrad/Kokoro-82M, released under Apache 2.0.

Dataset attribution

Training data was obtained from afkfatih/turkish-tts-combined-raw, published under CC BY-SA 3.0.