# Anka TTS

> Turkish text-to-speech and zero-shot voice cloning. Anka TTS is a 336M-parameter model on the F5-TTS architecture, trained for Turkish and shipped as the Python package `anka-tts` (Python 3.10+). It outputs 24 kHz mono audio and scores 1.73% WER on a 495-sentence Turkish test set, first of the eight systems measured.

Türkçe: Türkçe konuşma sentezi ve zero-shot ses klonlama. `pip install "anka-tts[tts]"` ile kurulur. Paketle gelen iki hazır sesle (`male`, `female`) referans vermeden konuşur; isteğe bağlı olarak 7-9 saniyelik bir referans kaydıyla kendi sesinizi klonlar.

Key facts:

- Install: `pip install anka-tts` gives text normalization and reference checks only, with no torch. `pip install "anka-tts[tts]"` adds synthesis and pulls torch (about 2 GB).
- Load: `AnkaTTS.from_pretrained("anka-tts/v0.1")`. The weights download from Hugging Face on the first call; the model runs on CUDA, MPS or CPU.
- Voices: built-in `male` (the default) and `female`, no reference needed. Register your own voice with `add_voice(name, wav_path, transcript)`.
- Text is normalized before synthesis: numbers, dates, times, money, abbreviations and Turkish suffixes are written out the way they are spoken.
- Licenses: code Apache-2.0; model weights and built-in voices CC-BY-NC-4.0 (non-commercial).

## Docs

- [Full reference](https://anka.krmkayabasi.com/llms-full.txt): install, API, voices, parameters, normalization, benchmark and licenses in one plain-text file
- [PyPI package](https://pypi.org/project/anka-tts/): the complete README, in English and Turkish
- [Model on Hugging Face](https://huggingface.co/krmkayabasi/Anka-TTS): weights, vocab and model card
- [Live demo (Hugging Face Space)](https://huggingface.co/spaces/krmkayabasi/Anka-TTS): browser demo with the built-in voices, zero-shot cloning and automatic reference transcription

## Website

- [Anka TTS home](https://anka.krmkayabasi.com/): overview, benchmark table and audio samples (Turkish, with an English toggle)

## Optional

- [F5-TTS](https://github.com/SWivid/F5-TTS): the architecture the model uses
- [Freya Eval TR](https://huggingface.co/datasets/freyavoice/freya-tr-eval): the evaluation set behind the benchmark