Loudness-ai/piper-voices-ranked

Open Piper TTS Voices Ranked

Every Piper text-to-speech voice that a business may use, scored for naturalness and clarity: 1,683 voices from 78 model files in 32 language variants, each with a listening sample, its licence and a pinned download link.

This is the dataset of the Loudness.ai technical report LA-RB-2026-003, Open Piper TTS Voices Ranked: Naturalness, Clarity and Licences (2026-10-15). Read the report, with an interactive explorer, at https://loudness.ai/research/piper-tts-voices-ranked/. Explore the voices in the browser: https://huggingface.co/spaces/Loudness-ai/piper-voice-explorer

What is in it

File	Rows	Contents
samples/ (the viewer)	1,683	One row per voice: sample (MP3), Overall, Natural and Clear scores, raw MOS and CER, rank within its language, quality, recorded sex of the reader, licence and download links
models.csv	78	One row per model file: language, quality, speakers, licence, training data, links pinned to the tested revision, size and checksum
languages.csv	32	Best and median voice per language variant
excluded.csv	102	Official files left out, with the reason and the licence their card states
speed.csv	8	CPU real-time factor per model on two threads

from datasets import load_dataset
voices = load_dataset("Loudness-ai/piper-voices-ranked", "voices", split="train")
best_english = voices.filter(lambda v: v["language"] == "English (United States)").sort("rank")[0]

import pandas as pd
models = pd.read_csv("hf://datasets/Loudness-ai/piper-voices-ranked/models.csv")

How each voice is scored

Every voice of a language reads the same three sentences (benchmark-texts.json; no digits or abbreviations).

• Natural = 100 × (MOS − 1) / 4, from the UTMOS22 "strong" predictor (SpeechMOS v1.2.0), averaged over the sentences.
• Clear = 100 × (1 − CER), from one Whisper large-v3 transcript (faster-whisper 1.2.1, int8, greedy, language forced).
• Overall = Natural / 2 + Clear / 2, plus 4 for voices recorded at 22.05 kHz (medium and high quality).

Compare voices within one language: the naturalness model was trained mainly on English and the recogniser is weaker in some languages. Scores repeat within a few points across machines (60 voices measured twice: r = 0.948, mean absolute difference 3.4 points). Sex is given only where the training dataset records the reader; it is never inferred from audio.

Best voice per language

Language	Voices	Best voice	Best	Median
English (United States)	1,158	Clean 100 · speaker 118	96.4	91.3
English (US/UK mix)	16	ManyVoice · speaker 2	95.0	89.2
English (Ireland)	1	Jenny	93.5	93.5
English (Great Britain)	4	Cori (medium)	92.4	90.1
Italian (Italy)	3	Serena (high)	89.6	86.3
Ukrainian (Ukraine)	7	Tetiana	88.1	84.8
Swedish (Sweden)	2	NST	86.4	86.3
Basque (Spain)	2	Maider	85.6	82.3
Dutch (Netherlands)	55	Alex	85.6	61.3
Icelandic (Iceland)	4	Steinn	85.3	83.2
Urdu (Pakistan)	1	Aegis Female	85.2	85.2
Spanish (Mexico)	2	Claude	85.0	79.5
Slovenian (Slovenia)	1	Artur	84.6	84.6
Farsi (Iran)	1	Gyro	84.2	84.2
Latvian (Latvia)	1	Aivars	82.8	82.8
Norwegian (Norway)	10	NVCC · speaker 6	82.0	77.8
German (Germany)	241	Karlsson	81.8	70.0
Chinese (China)	1	Huayan	79.6	79.6
Dutch (Belgium)	3	Nathalie	79.5	76.8
Czech (Czech Republic)	1	Kasandra	78.1	78.1
Greek (Greece)	2	Rapunzelina (medium)	77.0	72.2
Kazakh (Kazakhstan)	8	ISSAI · speaker 4	77.0	71.1
Finnish (Finland)	1	Harri	76.8	76.8
French (France)	129	Gilles	76.8	61.8
Catalan (Spain)	2	UPC Ona	76.6	76.2
Spanish (Spain)	3	MLS 9972	75.8	75.1
Portuguese (Brazil)	1	Edresson	74.1	74.1
Polish (Poland)	1	MLS 6892	68.8	68.8
Armenian (Armenia)	1	Gor	64.3	64.3
Telugu (India)	2	Padmavathi	56.8	54.1
Vietnamese (Vietnam)	1	25hours single	56.4	56.4
Nepali (Nepal)	18	Google · speaker 1	55.5	48.0

A community model trained from scratch, Clean 100, matches the best official voice (best speaker 96.4 against 96.2 for LibriTTS high) and runs four to six times faster on a CPU. Inside one model, speakers differ by up to 46 points, so speaker 0 is rarely the best choice.

Licences

Every voice here allows commercial use, with conditions given per row in use and licence: no conditions (51 voices, public domain or CC0), credit required (1,613, CC BY, CC BY-SA, Apache 2.0 or the Jenny licence), copyleft (2, GPL or AGPL) and unclear (17, check the source first). 102 official files were excluded: 76 because they 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. This is a summary, not legal advice.

The scores and tables are CC BY 4.0. Each audio sample is synthetic speech from the model named in its row and keeps that model's licence and credit requirements. See LICENSE.md.

Download a voice

model_url and config_url point to the original files at the exact version tested: the official Piper voices repository at revision c10ece1aade47bb51c153c893d14e5bf8e5b7117, or the community author's own hosting. Check them against md5 (catalog files) or sha256 (community files) in models.csv.

python3 -m piper -m en_US-libritts-high.onnx -s 0 -f speech.wav -- 'Hello'

Citation

@techreport{shehzad2026pipervoices,
  author = {Shehzad, Asad},
  title = {Open Piper TTS Voices Ranked: Naturalness, Clarity and Licences},
  institution = {Loudness.ai},
  number = {LA-RB-2026-003},
  year = {2026},
  url = {https://loudness.ai/research/piper-tts-voices-ranked/}
}

More from Loudness.ai

• Piper Voice Explorer: listen to, filter and download these voices
• Loudness.ai Text to Speech: free online text to speech
• loudness-tts (GitHub, PyPI): use these voices in Python or JavaScript; it picks the best voice for each language from this ranking
• Free Text to Speech in Your Browser: try the voices with word highlighting and subtitles
• AI Noise Remover SDK (GitHub): local noise removal, speech enhancement and loudness normalization
• Developer guides · Developer hub
• GitHub · Research & Benchmarks

Competing interests: the author develops Loudness.ai, whose text-to-speech tool offers a selection of these voices chosen with the same scores.