piper-liana-romanian

Liana — Romanian Piper voice for storytelling and long-form reading

Two fine-tuned Piper (VITS) voices — one medium (64 MB, real-time on CPU and phones) and one high (114 MB, best quality) — built for reading whole stories, articles and books aloud in Romanian. Non-autoregressive, CPU-only, no hallucinations, no drift.

Liana is the successor of the Lili voice from eduardem/piper-tts-romanian: the same narrator, retrained on a new purpose-recorded corpus and specifically polished on everything the old voices got wrong — long sentences, English loanwords, foreign names, and numbers.

Quick listen — „Punguța cu doi bani" by Ion Creangă, read end to end in one synthesis call (no chunking, no post-editing):

Liana medium (8:00): <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/stories/punguta-cu-doi-bani.mp3"></audio> Liana high (8:03): <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/stories/punguta-cu-doi-bani.mp3"></audio>

Voices

WER measured with Whisper large-v3 transcription against the input text, normalized before scoring (case, punctuation, digits written out, hyphenation variants) so that only real errors count. Raw unnormalized WER runs ~2–7 pp higher on the same audio; both numbers per test set are in the tables below. VITS synthesis is stochastic, so single measurements on small sets swing ±2–4 pp.

What makes Liana different

1. It does not fall apart on long input

The v1 voices collapse past ~200–300 characters in a single sentence (up to 90%+ WER — the old card documents it and tells you to chunk). Liana was trained on long spliced clips precisely to fix this. Same test, one unchunked sentence per synthesis call, five sentences per length bucket:

Lili's row was re-measured for this card, same session, same methodology, same patched dictionary — the collapse is the voice, not the setup.

It holds on real texts, fed whole — one synthesis call per article, Piper's own sentence splitting, nothing else:

And the two full stories in samples/…/stories/ — 4 and 8 minutes of continuous narration each — were generated exactly that way.

2. English loanwords and names, without tricks

Romanian everyday text is full of English — weekend, podcast, marketing, laptop, site, WhatsApp. The old voices depend on the patched dictionary for this (Lili v1: 10.4% on the loanword set stock, 3.9% with it); Liana was trained on it, on purpose-recorded loanword and name material. Same methodology, same dictionary: loanwords 3.3%, and on the English-names set 1.9% vs Lili's 4.0% — names are where dictionaries help least and training shows most.

"Filmele cu John Wayne se dădeau duminica după-amiaza, pe vremea aceea."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/names_john_wayne.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/names_john_wayne.wav"></audio>

3. Numbers read like a human reads them

Compound numerals come out in one breath — not "patruzeci … și … două" with a pause in the middle:

"A plătit două sute optzeci și șapte de lei pe patruzeci și două de bilete."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/numerale.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/numerale.wav"></audio>

"De luni se anunță temperaturi de treizeci și cinci de grade în oraș."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/luni_temp.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/luni_temp.wav"></audio>

4. Abbreviations, addresses, centuries — expanded before synthesis

The shipped ro_text.py front-end (single file, standard library only) normalizes diacritics, expands abbreviations to their spoken forms and knows Romanian grammar corner cases: bd. → bulevardul, nr. → numărul, 80 mp → 80 de metri pătrați, sec. XX → secolul al douăzecilea.

"Biroul s-a mutat pe bd. Unirii nr. 24, et. 3, ap. 12, într-un spațiu de 80 mp construit în sec. XX."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/frontend_adresa.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/frontend_adresa.wav"></audio>

5. A storytelling register

The corpus behind Liana was recorded in a projected, narrative read — the register of someone telling a story to a room, not dictating into a phone. It shows most on literature; hear the full Creangă stories below.

Audio samples

Full stories (Ion Creangă, public domain)

Each story is one synthesis call over the whole text.

Ursul păcălit de vulpe (4:13)

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/stories/ursul-pacalit-de-vulpe.mp3"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/stories/ursul-pacalit-de-vulpe.mp3"></audio>

Punguța cu doi bani (8:00)

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/stories/punguta-cu-doi-bani.mp3"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/stories/punguta-cu-doi-bani.mp3"></audio>

Short sentences

Same texts and file names as the old card, so you can compare the voices player for player.

"Copiii se jucau fericiți în grădina plină de flori colorate."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/sample_06.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/sample_06.wav"></audio>

"Într-o dimineață de toamnă, frunzele cădeau ușor pe aleea din parc."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/sample_04.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/sample_04.wav"></audio>

"Guvernul a anunțat un plan ambițios de investiții în infrastructură, care include construcția a trei autostrăzi și modernizarea rețelei feroviare din întreaga țară."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/sample_13.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/sample_13.wav"></audio>

Tongue twisters

"S-a suit capra pe piatră, piatra a crăpat în patru, crăpai-ar capul caprei negre în patru cum a crăpat și piatra în patru."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/tongue_twister_01.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/tongue_twister_01.wav"></audio>

"Ce-ntâmplare întâmplăreață s-a-ntâmplat în tâmplărie, un tâmplar din întâmplare s-a lovit cu tâmpla-n cap."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/tongue_twister_02.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/tongue_twister_02.wav"></audio>

English loanwords

"Weekendul acesta rămânem în oraș."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/loanwords/loanword_02.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/loanwords/loanword_02.wav"></audio>

"Echipa testează software-ul înainte de lansare."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/loanwords/loanword_03.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/loanwords/loanword_03.wav"></audio>

"Trimite-mi locația pe WhatsApp, ca să ajung mai repede."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/loanwords/loanword_23.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/loanwords/loanword_23.wav"></audio>

"Pe YouTube au publicat și subtitrarea în română."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/loanwords/loanword_27.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/loanwords/loanword_27.wav"></audio>

"A salvat traseul în Google Maps."

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/loanwords/loanword_19.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/loanwords/loanword_19.wav"></audio>

The old card's long text (1:49, one unchunked call)

The same 1,565-character paragraph the old card uses to demonstrate where the v1 voices break — three of its sentences are 292, 410 and 448 characters long. Compare with the players there.

Liana medium: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-medium/long_text.wav"></audio> Liana high: <audio controls src="https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/samples/liana-high/long_text.wav"></audio>

Quick start

pip install piper-tts

wget https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/voices/liana-medium/ro_RO-liana-medium.onnx
wget https://huggingface.co/eduardem/piper-liana-romanian/resolve/main/voices/liana-medium/ro_RO-liana-medium.onnx.json

echo "Bună ziua!" | piper -m ro_RO-liana-medium.onnx --output_file hello.wav

Keep each .onnx with its own .onnx.json — the configs are not interchangeable between voices.

Python, with the text front-end

import numpy as np
import soundfile as sf
from piper import PiperVoice, SynthesisConfig

from ro_text import prepare   # ships in this repo — normalize, expand

voice = PiperVoice.load("ro_RO-liana-medium.onnx")
sc = SynthesisConfig(
    length_scale=voice.config.length_scale,
    noise_scale=voice.config.noise_scale,
    noise_w_scale=voice.config.noise_w_scale,
)

text = open("poveste.txt", encoding="utf-8").read()

sentences = []
for piece in prepare(text):                       # no chunking — see below
    for chunk in voice.synthesize(piece, syn_config=sc):
        sentences.append(chunk.audio_float_array.astype(np.float32))

# Piper emits one chunk per sentence and no silence between them, so the pause
# a listener hears is whatever you join them with.
gap = np.zeros(int(0.2 * voice.config.sample_rate), dtype=np.float32)
parts = []
for i, sentence in enumerate(sentences):
    if i:
        parts.append(gap)
    parts.append(sentence)

audio = np.concatenate(parts)
sf.write("poveste.wav", audio, voice.config.sample_rate)

prepare() folds the Turkish cedilla ş/ţ into the Romanian comma-below ș/ț (web text carries the wrong pair constantly), NFC-normalizes, and expands abbreviations.

Do not chunk for Liana. Piper already splits at ., ? and ! and synthesizes one chunk per sentence, so chunking multi-sentence text buys nothing; all it adds is cutting a long sentence at its commas, which is precisely what Liana does not need — the table above is flat to 600 characters, and the stories were synthesized in one call each. The chunker is still in ro_text.py for the v1 voices, off by default: pass prepare(text, max_chars=140) to switch it on. Note that max_chars never joins sentences, so raising it does not give you longer pieces — the choice is chunking or no chunking.

The dictionary

The voices phonemize through espeak-ng's Romanian dictionary. This repo ships the patched ro_dict these voices are built around — 200+ added entries for English loanwords and proper names, respelled into Romanian phonemes (one respelling was further corrected during release listening; everything else is byte-identical to what the voices were trained and evaluated with). Installing it is a one-file swap into the espeak-ng-data directory of your Piper installation; the ro_extra source and build script are alongside, so it can be rebuilt or extended.

Without the patched dictionary the voices still work, but rare names fall back to espeak's Romanian letter-to-sound rules and come out mangled.

Known limitations

• A handful of rare foreign names and brands still render imperfectly — in our test battery: LinkedIn, Nietzsche, München, Zürich, Marseille, Auchan, ChatGPT, Kafka, Bilbao and a few similar. The dictionary source ships in this repo, so a better respelling is a one-line edit and a rebuild away — no retraining needed.
• VITS is stochastic: the same input renders slightly differently each run.
• The voices speak Romanian. English words in Romanian text are handled; English sentences are not.
• The literary register (19th-century prose, dialogue interjections) is measurably harder than modern text — see the Creangă row in the article table.

Training

Both voices continue the Lili lineage: they warm-start from the released v1 weights and are retrained on a new ~25-hour corpus recorded by the same narrator specifically for this project — long spliced clips (up to ~55 s) for endurance, wide-coverage sentences for vocabulary, and a storytelling register throughout. Two low-learning-rate polish rounds followed, on a ~1,000-clip curated booster recorded to fix what the evaluation battery and human listening found: English loanwords and names, compound numerals read in one breath, and specific phonetic gaps. Checkpoint selection at every stage was done by transcription-WER evaluation and human listening, never by training loss.

• Framework: piper1-gpl v1.4.1
• Evaluation: Whisper large-v3 transcription; a 444-sentence Romanian test battery (names, loanwords, numbers, abbreviations, phoneme-grapheme traps, prosody) plus long-form article and sentence-length stress suites
• The training corpus is not published.

Files

voices/
├── liana-medium/
│   ├── ro_RO-liana-medium.onnx           # 64 MB
│   └── ro_RO-liana-medium.onnx.json
└── liana-high/
    ├── ro_RO-liana-high.onnx             # 114 MB
    └── ro_RO-liana-high.onnx.json
ro_text.py                                # text front-end: normalize + expand
espeak/
├── ro_dict                               # prebuilt patched dictionary (drop-in)
├── ro_extra                              # dictionary source
├── build_ro_dict.sh                      # rebuilds ro_dict from source
└── espeak-ng.pin                         # the upstream commit the build pins to
samples/
├── liana-medium/                         # everything embedded above, plus more
│   ├── loanwords/
│   └── stories/                          # full stories, WAV + MP3
└── liana-high/
texts/                                    # the story texts (public domain, provenance inside)