ahenk-turkish-tts
Copyright 2026 Özgür Güler

This product's own code, documentation, test sets (TR-Zor, TR-Soru, dev_numbers10, contest) and
text bank (tts/voxcpm/textbank) are licensed under the Apache License, Version 2.0 (see LICENSE).

-------------------------------------------------------------------------------
Third-party material included in this repository
-------------------------------------------------------------------------------

The files below contain text from third-party datasets (reference transcripts used to score
models). They are redistributed under their own licences, not under Apache-2.0:

- benchmarks/asr/results/fleurs_tr_test/*.jsonl ("ref" fields)
  FLEURS, Google, https://huggingface.co/datasets/google/fleurs, CC BY 4.0.
  Changes: Turkish test split only; model transcripts added next to the references.

- benchmarks/asr/results/mediaspeech_tr/*.jsonl ("ref" fields)
  MediaSpeech (Turkish), NTR Lab, https://huggingface.co/datasets/ymoslem/MediaSpeech, CC BY 4.0.
  Changes: a seeded 500-row sample of the Turkish subset; model transcripts added next to the references.

- benchmarks/tts/sets/freya_sample40.jsonl and benchmarks/tts/results/freya_sample40/
  Freya-TR-Eval, Freya Voice, https://huggingface.co/datasets/freyavoice/freya-tr-eval, CC BY 4.0.
  Changes: a seeded sample of 40 rows (benchmarks/tts/make_dev_sets.py); scores added.

-------------------------------------------------------------------------------
Datasets and models used (not included in this repository)
-------------------------------------------------------------------------------

Released model "Ahenk" (https://huggingface.co/ozgurguler/ahenk-tts-tr), a LoRA adapter
trained with the recipe in tts/voxcpm. Its weights are CC BY 4.0 and its model card carries the
full attribution. It was trained with:

- Alania Turkish Synthetic Speech by PatientDesk AI
  (https://huggingface.co/datasets/cloud0day3/alania-synthetic-speech-tr), licensed under CC BY 4.0.
  Only the "cc-by" configuration is used (never "cc-by-sa"). Includes Turkish web text from
  FineWeb-2 (https://huggingface.co/datasets/HuggingFaceFW/fineweb-2), used under ODC-By 1.0 and
  subject to Common Crawl's terms of use.
- Alania Turkish Domain Text by PatientDesk AI
  (https://huggingface.co/datasets/cloud0day3/alania-domain-text-tr), licensed under CC BY 4.0.
- VoxCPM2, OpenBMB (https://huggingface.co/openbmb/VoxCPM2), Apache-2.0: base model, voice design
  and the synthetic training audio.

Earlier experiments and evaluation:

- Antalia (Patientdesk.ai), Antalia Turkish single-speaker scripted speech,
  https://huggingface.co/datasets/cloud0day3/antalia-voice-corpus, CC BY 4.0. Used for the
  stopped Piper experiment (tts/) and the first VoxCPM2 LoRA; its 20 held-out sentences are an
  evaluation set. Not in the training data of the released Ahenk version. The speaker is a
  real, consenting voice actor: do not use anything trained on it to impersonate her or any real
  person, and do not present its output as a genuine recording.
- EMA Lightning (https://huggingface.co/canberkkkkkk/ema-lightning), Apache-2.0: the voice of
  `trtts` (downloaded at run time, not included) and the starting point of the on-device student
  experiments in tts/small.
- Antalia-2 Mini (https://huggingface.co/cloud0day3/antalia-mini), Apache-2.0: benchmarked only.
- Other models compared in benchmarks/ (Whisper, Vosk, Piper, MMS-TTS, pocket-tts-turkish,
  qwen3-tts-tiny-tr, normalizer-tr, Trendyol-TTS and others) are used under their own licences and
  are not redistributed here.

-------------------------------------------------------------------------------
Synthetic speech
-------------------------------------------------------------------------------

Speech produced with the models and recipes in this repository is synthetic. Disclose that it is
AI-generated wherever a listener could be misled (EU AI Act, Article 50), and do not use it to
impersonate any real person.
