---
language: tr
license: mit
tags:
- piper-tts
- text-to-speech
- turkish
- onnx
- quantized
thumbnail: null
---

# Piper TTS - Turkish (Cem)

Single-speaker Turkish TTS model (speaker: **Cem**, 4490 utterances).

- **Sample rate:** 24000 Hz
- **Phonemes:** 256 (espeak-ng/tr)
- **Speakers:** 1 (Cem)
- **Architecture:** Piper custom hybrid (resblock=1, upsample_initial_channel=512)
- **Inference params:** 21.0M
- **Training:** 437 epochs, ~1.77M steps

## Available Models

| Variant | Size | Type |
|---------|------|------|
| [`last.onnx`](last.onnx) | 84.9 MB | FP32 (original) |
| [`last.q8.onnx`](last.q8.onnx) | 30.8 MB | INT8 dynamic quantization (recommended for CPU/RPi) |

**Q8 (INT8)** gives ~64% size reduction with minimal quality loss. Ideal for Raspberry Pi and CPU inference.

## Sample Audio

Cem voice reading a Turkish finance/economy paragraph (~24s, 410 chars):

Epoch 426:
<p><audio controls src="https://huggingface.co/dcx514ai/piper_tts_turkish_xlarge/resolve/main/sample_tr.wav"></audio></p>

Epoch 450~
<p><audio controls src="https://huggingface.co/dcx514ai/piper_tts_turkish_xlarge/resolve/main/epoch450~.wav"></audio></p>

Inference CPU - intel 14600K - 0.4 avg_rtf


## Usage

```bash
# With piper CLI
echo "Merhaba dünya" | piper --model last.q8.onnx --output-raw | aplay -r 24000 -f S16_LE -t raw
```

```python
import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("last.q8.onnx")
dummy = {
    "input": np.random.randint(0, 100, (1, 50), dtype=np.int64),
    "input_lengths": np.array([50], dtype=np.int64),
    "scales": np.array([0.667, 1.0, 0.8], dtype=np.float32),
}
audio = session.run(["output"], dummy)[0]
```

Or with [piper-phonemize](https://github.com/rhasspy/piper-phonemize) + [piper-tts](https://github.com/rhasspy/piper).