Open Piper TTS Voices Ranked: Naturalness, Clarity and Licences

What we tested

The official Piper voice catalog (revision c10ece1, as tested in October 2026) holds 177 voice files in 53 languages. Before testing, we read every voice card and its training data licence. 102 files were set aside: 76 because their models were fine-tuned from the Lessac voice, whose recordings are licensed for non-commercial use, and 26 because their own licence is non-commercial or research-only. That left 75 official files in 27 languages that a business can use.

We added three community models trained from scratch by Bryce Beattie, who also made several official voices: Clean 100 (245 speakers), ManyVoice (16 speakers) and Jenny (Irish English). His Norman file is identical to the official one, so it is listed once. Then we tested every speaker of every multi-speaker model: 1,683 voices from 78 model files in 32 language variants.

How each voice is scored

Every voice of a language reads the same three sentences: a short self-introduction, a sentence about a Monday meeting and a question about the nearest train station. They contain no digits or abbreviations, so text handling does not affect the result.

Natural (0–100) is the predicted naturalness from UTMOS22, a model trained on human ratings of synthetic speech, averaged over the three sentences and rescaled from its 1–5 scale. Clear (0–100) is 100 minus the character error rate of a Whisper large-v3 transcript against the text the voice was given. Overall is half of each, plus 4 points for voices recorded at 22 kHz, whose extra bandwidth the 16 kHz test clips do not show.

Rankings compare voices within one language only. The naturalness model learned mostly from English, and the recogniser is weaker in some languages, so a score of 80 in Telugu and 80 in English do not mean the same thing. Scores are reproducible to within a few points: 60 voices measured in both rounds, on different machines, agreed with a correlation of 0.948 and a mean difference of 3.4 points. Natural and Clear agree in general (correlation 0.80), but not voice by voice, which is why we report both.

Explore and download every voice

Filter by language, gender, licence, quality and score, listen to each voice and download its model. Gender appears only where the training dataset records the reader; it is never guessed from the sound. Downloads come straight from the original source at the exact version we tested: the official Piper repository on Hugging Face, or Bryce Beattie’s own hosting for the community models.

The best voice in each language

English is far ahead: all 793 voices that score 90 or more are English. Most other languages have between one and three usable voices, and several (Nepali, Telugu, Vietnamese and Armenian) score low partly because the recogniser itself struggles there; a native listener should judge them.

What stands out

A community voice matches the best official one, and runs four to six times faster. The best Clean 100 speakers score 96.4, beside 96.2 for the best speaker of the official high-quality LibriTTS model. Clean 100 is also more even: its median speaker scores 93.2 against 90.8, and its 245 speakers span 9.3 points against 17.6 for LibriTTS’s 904. Being a medium-quality model, it synthesised speech in 0.091 of real time on two CPU threads of our test laptop, against 0.568 for LibriTTS high, and on our server 0.193 against 0.834.

The speaker you pick matters more than the model. Inside one model, speakers differ by up to 46.3 points (French MLS: 24.4 to 70.7) and 31.0 points in German MLS. Speaker 0, the default in most tools, is rarely the best.

Female and male voices score alike. Among LibriTTS’s readers, women average 90.8 and men 90.1.

Most of the catalog is not safe for commercial use. 102 of 177 official files (58%) were excluded, most because they inherit the Lessac voice’s non-commercial licence through fine-tuning. Their own cards often list a permissive dataset licence, which is easy to misread.

Speed on an ordinary CPU

Speed depends on the model’s quality tier, not on the speaker. We timed each model we hold locally on two CPU threads, as a small server would run it. The real-time factor is seconds of computing per second of speech: 0.1 means ten times faster than real time. On a 12-core AMD EPYC virtual server the same settings measured 0.17–0.19 for medium voices and 0.83 for LibriTTS high, about 1.5 to 2 times slower than the laptop.

Licences: what you may do

Every voice in this study allows commercial use, with the conditions shown on its badge. No conditions (51 voices): public domain or CC0. Credit required (1,613): CC BY, CC BY-SA or Apache 2.0; show a short credit line where your users can see it. CC BY-SA also asks that adaptations of the voice keep the same licence. Copyleft (2): GPL or AGPL model licences, which carry software-style sharing duties. Unclear (17): the card points to a web page or states nothing; check the source before using them commercially. The Jenny licence also requires the voice to be called “Jenny” in software that generates audio.

This is a summary, not legal advice. The explorer shows each model’s licence, its training data and a credit line you can copy.

Why 102 official voices are left out

Fine-tuning starts a new voice from an existing model, so the new voice inherits what that model learned. 76 official files were fine-tuned from the Lessac voice, directly or through another voice, and the Lessac recordings are licensed for non-commercial use. We excluded them as a precaution even when their own dataset is permissive; the legal position of such models is unsettled. The other 26 files carry a non-commercial or research-only licence themselves.

Limitations

All scores are automatic; no human panel rated these voices yet. The naturalness model was trained mostly on English, and the three test sentences per language were written for this benchmark, not by native speakers. Each sentence was synthesised once, and Piper’s output varies a little from take to take. Round one ran on a server and round two on a laptop; the two agree closely but not exactly. Gender comes only from dataset records, so many single-speaker voices show “not recorded”. Licence summaries reflect each voice card at the tested revision.

Data and reproduction

The study package contains every voice’s scores (CSV and JSON), each model’s licence, source, size and checksum, the test sentences, the listening samples and the scripts used for synthesis, scoring, analysis and speed. Use Piper 1.8, faster-whisper 1.2.1 with Whisper large-v3, and the SpeechMOS port of UTMOS22 to repeat the measurements.

The same data is on Hugging Face as a dataset with an audio player for every voice, Loudness-ai/piper-voices-ranked, and the voice explorer also runs as a Hugging Face Space.

To use the ranking in your own software, loudness-tts, our free and open-source text-to-speech SDK for Python and JavaScript, picks the best voice for each language from this study and shows its licence and credit line. Try it with Free Text to Speech in Your Browser, or follow the installation guide.

Competing interests and funding

The author develops Loudness.ai, whose Text to Speech tool offers a selection of these voices chosen with the same scores. No external funding was received.

References

Hansen, M. and the Open Home Foundation. Piper; Piper voices, revision c10ece1. Beattie, B. Community Piper voices.

Saeki, T. et al. (2022). UTMOS: UTokyo-SaruLab system for VoiceMOS Challenge 2022. Interspeech 2022. Radford, A. et al. (2023). Robust speech recognition via large-scale weak supervision. ICML 2023. Koizumi, Y. et al. (2023). LibriTTS-R. Interspeech 2023. Pratap, V. et al. (2020). MLS: a large-scale multilingual dataset for speech research. Interspeech 2020.

Paper and study files

Cite this report

Shehzad, Asad (2026). Open Piper TTS Voices Ranked: Naturalness, Clarity and Licences. Loudness.ai, LA-RB-2026-003, version 1.0. https://loudness.ai/research/piper-tts-voices-ranked/

BibTeX · RIS · Research RSS feed