yazgı-tts-1

A Turkish text-to-speech model with 8 built-in voices that speaks clean, accurate and natural Turkish. It reads numbers, prices, dates, codes and addresses correctly, and runs about 6x faster than real time on a laptop CPU. On Freya-TR-Eval it is one of the most accurate and natural open Turkish models.

Samples

Benchmarks

Freya-TR-Eval (495 sentences), scored with the Freya-TR-Eval re-run by PatientDesk AI: 3 seeded takes per system, faster-whisper large-v3 on audio band-limited to 8 kHz, corpus WER per take. Naturalness is UTMOS22, out of 5.

¹ From the EMA Lightning model card.

² From the Antalia-2 Mini model card.

Other rows are from the Freya-TR-Eval re-run.

Speed: real-time factor 0.17 on a laptop CPU (Apple M5, 4 threads), about 6x faster than real time. With streaming, the first audio is ready after 0.27 s on the same CPU.

Quick start

pip install "yazgi-tts[hub] @ https://huggingface.co/erkamk/yazgi-tts-1/resolve/main/dist/yazgi_tts-0.1.0-py3-none-any.whl"
python -m yazgi_tts "Siparişiniz 14 Ekim'de saat 15:30'da kargoya verildi." -o out.wav --voice male-01

from yazgi_tts import YazgiTTS, RATE
import soundfile as sf

tts = YazgiTTS()
audio = tts.say("Toplam tutar 1.299,90 TL.", voice="female-01")
sf.write("out.wav", audio, RATE)

# optional: your own readings for names the front end reads wrongly
tts = YazgiTTS(pronunciations={"Xiaomi": "şaomi"})

# streaming: play audio while the rest is still being made
for chunk in tts.stream("Merhaba! Siparişiniz hazırlanıyor.", voice="male-01"):
    ...  # each chunk is 24 kHz float32 audio

HTTP API: python -m yazgi_tts.serve --port 8000, then POST /tts {"text": "...", "voice": "male-01"} for a WAV file, or POST /tts/stream with the same body for raw 16-bit PCM at 24 kHz, sent while it is made.

Model details

• Codec: 24 kHz audio as 12.5 frames per second, each frame 16 codes from 2,048 (2.2 kbps). Decoder 25M; the 72M encoder is only needed to clone a new voice.
• TTS model (56M): an autoregressive talker writes the first code of each frame; a small code predictor fills the other 15. The voice comes from a short reference and a speaker embedding.
• Text front end: normalizer-tr 0.7.0 plus our rules for initialisms, codes and units. Long text is split into whole sentences (up to 300 characters) and decoded in one pass.
• 81M parameters for speech in total.

Training

1. Codec: 559k steps on Alania and about 600 h of real Turkish speech.
2. Supervised training: about 120k steps on 2,485 h of Alania (1.65M clips, 2,706 voices).
3. Speaker-embedding fine-tune: 15k steps.
4. Reinforcement learning (GRPO): 125 steps, rewarding reading accuracy, voice similarity and naturalness.

Data and licenses

Held out of all training: FLEURS dev and test, 1 in 40 Common Voice speakers, and Freya-TR-Eval.

Limitations

• Little emotion. It reads clearly and calmly but does not show emotion such as excitement or sympathy.
• Voice cloning is weak. tts.clone() builds a voice from 3-6 s of speech, but it only sounds close for voices like the built-in ones (similarity about 0.27 on real speakers). Use the built-in voices.
• Turkish only. Foreign names, brand names and words are read with Turkish letter sounds; pass pronunciations to fix the ones you need.

Attribution

YODAS2 videos are licensed by their uploaders; the video ids are listed in ATTRIBUTION.md. Alania's speech was generated with VoxCPM2, and part of its text comes from FineWeb-2 (ODC-By 1.0).

License

Apache-2.0 for the code and the weights; see LICENSE. Credits for the training data are in NOTICE and ATTRIBUTION.md.