# Server-side Piper voices The binary voice files are provisioned during deployment and are intentionally not committed below `tts/voices/`. ## Locked source - Repository: `WitoldG/polish_piper_models` - Revision: `ed4cda3cd323636873ffc563bc692f50e9dca9fa` - Source metadata: - License declared by the source repository: MIT The deployed files and their pinned download URLs are listed in `voices.sources.tsv`. `voices.sha256` contains the checksums verified against the existing desktop models in `resources/piper/` and the source repository's LFS metadata. | Product voice | Piper model | | --- | --- | | Natan PL | `pl_PL-jarvis_wg_glos-medium` | | Jowita PL | `pl_PL-justyna_wg_glos-medium` | Run `scripts/provision-tts-voices.sh` before starting the stack. Run `scripts/verify-tts-service.sh` after the TTS container becomes healthy; it performs a real WAV synthesis with both locked voices. ## Additional voices All voices are synthetic text-to-speech voices. Model weights are unmodified. | Voice | License | Source | | --- | --- | --- | | Joe EN | CC0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Thorsten DE | CC0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Anna HU | CC0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Faber BR | CC0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Dave ES | CC0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Vais VI | CC-BY-4.0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Edon AL | CC0 | https://huggingface.co/rhasspy/piper-voices/tree/c10ece1aade47bb51c153c893d14e5bf8e5b7117 | | Antalia TR | CC BY 4.0 | https://huggingface.co/ozgurguler/piper-tr-antalia-medium/tree/4e510fcb99a1b9e63238e6d9a36686173d5d540c | Full source model cards are included next to the models. CC0: https://creativecommons.org/publicdomain/zero/1.0/ . CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Vietnamese: VAIS-1000 Vietnamese Speech Synthesis Corpus, https://ieee-dataport.org/documents/vais-1000-vietnamese-speech-synthesis-corpus ; Piper model distributed by Michael Hansen / rhasspy, CC BY 4.0. Turkish: Antalia (Patientdesk.ai), Antalia Turkish single-speaker scripted speech, https://huggingface.co/datasets/cloud0day3/antalia-voice-corpus (CC BY 4.0). Piper model by ozgurguler, CC BY 4.0. Source training changes: resampling 24 kHz to 22.05 kHz, transcript normalization, mismatched clips removed and 20 clips held out. We do not modify weights. Without trnorm, numbers/dates are pronounced with espeak-ng rules. # Model card for joe (medium) * Language: en_US (English, United States) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://github.com/OHF-Voice/voice-datasets * License: CC0 ## Training Finetuned from U.S. English lessac voice (medium quality). # Model card for thorsten (medium) * Language: de_DE (German, Germany) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://github.com/thorstenMueller/Thorsten-Voice * License: CC0 ## Training Finetuned from U.S. English lessac voice (medium quality). # Model card for anna (medium) * Language: hu_HU (Hungarian, Hungary) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://github.com/OHF-Voice/voice-datasets * License: CC0 ## Training Finetuned from U.S. English lessac voice (medium quality). # Model card for faber (medium) * Language: pt_BR (Portuguese, Brazil) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://github.com/OHF-Voice/voice-datasets * License: CC0 ## Training Finetuned from U.S. English lessac voice (medium quality). # Model card for davefx (medium) * Language: es_ES (Spanish, Spain) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://github.com/OHF-Voice/voice-datasets * License: CC0 ## Training Finetuned from U.S. English lessac voice (medium quality). # Model card for vais1000 (medium) * Language: vi_VN (Vietnamese, Vietnam) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://ieee-dataport.org/documents/vais-1000-vietnamese-speech-synthesis-corpus * License: https://creativecommons.org/licenses/by/4.0/ ## Training Finetuned from U.S. English lessac voice (medium quality). # Model card for edon (medium) * Language: sq_AL (Albanian/Shqip, Albanian) * Speakers: 1 * Quality: medium * Samplerate: 22,050Hz ## Dataset * URL: https://huggingface.co/edonseki/folsh.ai * License: CC0 ## Training Finetuned from U.S. English lessac voice (medium quality). --- language: - tr license: cc-by-4.0 library_name: piper pipeline_tag: text-to-speech tags: - piper - vits - turkish - synthetic-voice datasets: - cloud0day3/antalia-voice-corpus --- # tr_TR-antalia-medium — Türkçe Piper sesi (Antalia) > **Bu bir sentetik (yapay) sestir.** Gerçek bir seslendirme sanatçısının izinli kayıtlarıyla > eğitildi. Dinleyicinin yanılabileceği her yerde sesin yapay olduğunu belirtin; bu sesi > sanatçıyı ya da herhangi bir gerçek kişiyi taklit etmek için **asla** kullanmayın. > > **This is a synthetic voice.** It was trained on consented recordings of a real voice actor. > Disclose that it is synthetic wherever a listener could be misled, and **never** use it to > impersonate the actor or any real person. | | | |---|---| | Model | Piper VITS "medium" (~15.7 M generator parameters at inference), 22,050 Hz, mono | | Dil / Language | Türkçe (`tr_TR`), espeak-ng `tr` fonemleri | | Eğitim / Training | LJSpeech "medium" kontrol noktasından ince ayar (fine-tuned from LJSpeech medium): 426 epoch, 13632 batch (`last.ckpt`) | | Dışa aktarma / Exported | 2026-10-07, ONNX 60.6 MB | | Lisans / Licence | **CC-BY-4.0** (ağırlıklar / weights) | | Ön uç / Front-end | metni **mutlaka** `trnorm.normalize` ile verin / always feed text through `trnorm.normalize` | ## Kullanım / Usage ```bash pip install piper-tts trnorm # trnorm: https://github.com/ozguradmin/turkish-tts python -c "from trnorm import normalize; print(normalize('Randevunuz 14.10.2026 saat 09:30\'da.'))" \ | piper -m tr_TR-antalia-medium.onnx -f cikti.wav ``` Home Assistant: copy `tr_TR-antalia-medium.onnx` and `tr_TR-antalia-medium.onnx.json` into the Piper add-on's `/share/piper/` (or the `--data-dir` of `wyoming-piper`), restart, and pick the voice. Home Assistant does not run trnorm; numbers, dates and abbreviations are then read by espeak-ng's own rules. ## Eğitim verisi ve atıf / Training data and attribution - **Antalia (Patientdesk.ai)** — *Antalia Turkish single-speaker scripted speech*, CC BY 4.0, https://huggingface.co/datasets/cloud0day3/antalia-voice-corpus — one professional female voice actor who recorded under a commercial agreement and signed an addendum permitting public redistribution. Changes made: resampled 24 kHz → 22.05 kHz; transcripts converted to spoken form with trnorm; clips whose text disagreed with the corpus' ASR transcript removed; 20 clips held out. - **Base checkpoint:** `lj-med_1000` (Piper "medium", trained from scratch for 1000 epochs by Bryce Beattie) from `rhasspy/piper-checkpoints` (MIT), trained on **LJ Speech** (Keith Ito & Linda Johnson, public domain). - Training code: [piper1-gpl](https://github.com/OHF-Voice/piper1-gpl) (GPL-3.0) at a pinned commit; recipe and front-end: [turkish-tts](https://github.com/ozguradmin/turkish-tts) (Apache-2.0). Bu ağırlıkları kullanırken yukarıdaki Antalia atfını koruyun (CC-BY-4.0 şartı). Keep the Antalia attribution above when you use or redistribute these weights (CC-BY-4.0). ## Etik ve kabul edilebilir kullanım / Ethics and acceptable use Veri setinin sahipleri, konuşmacının izninin şu anlayışla verildiğini belirtiyor ve buna uyulmasını rica ediyor; biz de aynı ricayı bu model için tekrarlıyoruz: - Bu modeli **konuşmacıyı veya herhangi bir gerçek kişiyi taklit etmek**, birine söylemediği bir şeyi söyletmek için kullanmayın. - Üretilen sesi **gerçek bir insan kaydı gibi sunmayın**; dinleyicinin yanılabileceği yerlerde (telefon, IVR, duyuru, video) sentetik olduğunu açıkça belirtin. - **Dolandırıcılık, sosyal mühendislik, siyasi otomatik arama** için kullanmayın. The dataset owners state that the speaker consented on this understanding and ask users to honour it; we ask the same for this model: no impersonation of the speaker or any real person, no passing synthetic speech off as a genuine recording (disclose it where a listener could be misled), no fraud, scams, social engineering or political robocalls. (Under CC-BY-4.0 these are requests, not licence conditions; the Antalia 1 model of the dataset owners is Open RAIL-M, which makes them binding.) ## Sınırlamalar / Limitations - VITS + espeak-ng: anlaşılır ve çok hızlı, ama büyük modeller (EMA Lightning, pocket-tts-turkish) kadar doğal değil; vurgu ve tonlama espeak-ng'nin Türkçe kurallarına bağlı. - Yalnızca okuma konuşması, tek konuşmacı, çağrı merkezi / asistan üslubu ağırlıklı. - Metin `trnorm` ile normalize edilmeden verilirse sayıları espeak-ng okur (farklı okunuşlar). - Bilinen okunuş farkları: seslendirmen `TL`'yi "Türk lirası", kodlardaki tireyi "tire" diye okudu; trnorm "lira" yazıyor ve tireyi atlıyor (bkz. turkish-tts `tts/DATA_REPORT.md`). - espeak-ng (GPL-3.0) fonemleştirme için gerekir; piper-tts de GPL-3.0'dır. Ağırlıklar CC-BY-4.0; piper'ı ayrı bir servis olarak (Wyoming/HTTP) çalıştırmak ticari kullanımda en sorunsuz yoldur.