Ahenk: Turkish text-to-speech

English · Türkçe

Open Turkish speech tools: Ahenk (large model, GPU), Ahenk Cep (8.6M-parameter small model, CPU), trnorm (Turkish text normalisation) and the TR-Zor / TR-Soru test sets.

Synthetic voice. Ahenk and Ahenk Cep speak in one designed voice (f022); it is not a real person's voice or recording. Say that it is AI-generated when you share what you make (EU AI Act Art. 50).

Try it in your browser · Ahenk Okur: listen to documents · Ahenk model card · Ahenk Cep model card

GitHub does not play audio in the page: the links open the WAV files on Hugging Face. The model cards have the same samples with players. You can try Ahenk Cep in your browser, no install: ozguradmin.github.io/ahenk-turkish-tts.

How they were picked: the plain, money, laughing and emotion clips were chosen by the listener in blind listening. The question pair and the long sentence were chosen by automatic scores (word-perfect, highest UTMOS). All are existing evaluation clips, not rendered again.

The browser demo and Ahenk Okur process text with trnorm-js. It gave the same result as the Python version on 742,716 checks, so the browser no longer has to download Python: the first visit fell from 27 MB to 21.4 MB and the page is ready about twice as fast. The Python version stays as a fallback.

Ahenk Cep, on a CPU:

pip install torch ema-lightning==1.1.2 "git+https://github.com/ozguradmin/ahenk-turkish-tts"

import sys
import soundfile as sf
from huggingface_hub import snapshot_download

d = snapshot_download("ozgurguler/ahenk-cep-tr", allow_patterns=["*.pt", "ahenk_cep.py"])
sys.path.insert(0, d)
from ahenk_cep import AhenkCep

tts = AhenkCep(f"{d}/student.pt", f"{d}/decoder.pt")
audio, sr = tts.say("Randevunuz 14.10.2026 saat 09:30'da; ücret 1.250 TL.")
sf.write("randevu.wav", audio, sr)

Setup and emotions for Ahenk (GPU, VoxCPM2) are in its model card. Text normalisation only:

trnorm "Saat 09:30'da 3. kata gelin, ücret 1.250,50 TL."
# Saat dokuz otuzda üçüncü kata gelin, ücret bin iki yüz elli lira elli kuruş.

trnorm's rules and examples: docs/TRNORM.md.

TR-Zor: 155 hard sentences. Held-out: 20 long sentences (12-19 s) of the Antalia dataset, not in the training data of any of our models. Judge: faster-whisper large-v3-turbo; reference and output normalised with trnorm.normalize_for_eval. WER %, lower is better; in brackets, the number of sentences without a single word error. UTMOS22 predicts naturalness (1-5, higher is better).

• Ahenk and Trendyol-TTS were measured on an A100, the small models on a laptop GPU, with the same judge. Scored again on the second setup, Ahenk's clips came out the same within 0.01.
• Where we lose: Trendyol-TTS is ahead on predicted naturalness, most on long sentences (3.73 vs 3.14); it has a lower WER on the held-out sentences (0.68 vs 0.82) and is faster (RTF 1.37 vs 2.28 on an A100). Trendyol speaks in its own voice; part of the difference may come from the voice. On CPU, Ahenk Cep is slower than EMA and Antalia (RTF 0.20 vs 0.13 and 0.08 on 2 threads).
• trnorm's effect: EMA Lightning's TR-Zor error is 14.0 % with its own front-end and 1.9 % with trnorm (this row was measured on CPU with an int8 judge, do not compare it directly with the table above; details).
• These are automatic metrics. No MOS study was done.

Blind listening. All done with one native listener (the project owner), choosing the best clip for each sentence; it is a signal, not a statistical result.

• 10 sentences, four systems in random order, equal loudness: Ahenk 8, Ahenk Cep 2 (both questions), EMA Lightning 0, Antalia-2 Mini 0.
• Language of the emotion instruction, 20 decisions: for pure emotion the English instruction 6/6; for sentences with a word starting with "c" in their first two words, the Turkish instruction 10/10. instruction_pick in Ahenk works from this result.

• Questions that should rise at the end do not: questions without "mı", such as "Yarın geliyorsun?", end like a statement, in both models. Questions built with "mı" are right.
• "c" after an English instruction: in Ahenk a leading "c" may be read like /k/, /s/ or /z/ ("Cam" → "Kam"). instruction_pick switches to a Turkish instruction; if needed, use prefer="tr".
• "Phone-like" timbre: in some of Ahenk's emotional clips the voice is bright and crackly (7 of 20 clips in a probe with English instructions). Try another seed.
• Breath in pauses: stays in Ahenk; pause_clean in the package turns it down.
• Ahenk Cep's styles are weak: "sad" is only 1.9 semitones lower and 8 % slower; no laughter. Not yet distilled for phone speed.
• Ahenk needs a GPU: RTF 2.28 at 16 steps on an A100, slower than real time.
• One voice, Turkish only. trnorm leaves foreign names that are not in its lexicon as written.

• Do not present generated speech as a real human recording; say that it is synthetic and, if you can, add a machine-readable mark (EU AI Act Art. 50). The models have no watermark; AudioSeal can be added.
• Do not impersonate anyone. VoxCPM2 can clone voices; do not clone anyone's voice without their consent.
• Do not present generated anger, sadness or warmth as a real person's reaction to put pressure on a listener.
• Do not use it for fraud, social engineering or robocalls.

• Code, docs, the TR-Zor and TR-Soru sets: Apache-2.0. Third-party reference texts in the repo come with their own licences: NOTICE.
• Ahenk and Ahenk Cep weights are CC BY 4.0. Credit is required when you use or redistribute them. Write: "Ahenk (https://github.com/ozguradmin/ahenk-turkish-tts), CC BY 4.0". Also keep the data attributions from the model cards: Alania Turkish Synthetic Speech by PatientDesk AI and Alania Turkish Domain Text by PatientDesk AI (CC BY 4.0).
• Base models: VoxCPM2 (OpenBMB) and EMA Lightning, both Apache-2.0.
• Academic citation: CITATION.cff.
• Not required, but if you use this in a project, we would be glad of a link back or a quick note (destek@ozgurguler.tech).

• Make Ahenk Cep run in real time on phones (few-step distillation)
• More teacher data for the small model, stronger emotions and question intonation
• A demo page that runs in the browser
• A small, commercially usable Turkish speech-recognition (STT) model

Contributing: open an issue; sentences that trnorm reads wrong help most (input, expected reading, the output you got). python -m pytest -q must pass.

Contact: destek@ozgurguler.tech