tts-pl-piper-v2 — Polish Piper (VITS) fine-tuned on 8 Wolne Lektury readers

Control model for hvsr-robotics/tts-pl-matcha-v2: the same 8 readers and training clips, VITS architecture via piper1-gpl, exported to ONNX. Fastest option on CPU (≈ 60 ms per sentence on an Apple M-series core, 22.05 kHz), lower naturalness and intelligibility than the Matcha model (UTMOS 2.7–3.3, WER 7–22 %; VITS drops words for two readers).

Files

• export/piper_v2_target/pl_PL-tts-pl.onnx + .onnx.json — standard Piper voice files (piper -m pl_PL-tts-pl.onnx --speaker 7).
• UI and synth_piper.py (batch synthesis with the piper-tts package): tts-pl-playground.
• samples/, ATTRIBUTION.md, LICENSE, RECIPE.md (section 4 for Piper).

Speakers: 0 Jan Staszczyk, 1 Radosław Krzyżowski, 2 Wojciech Masiak, 3 Bartosz Bielenia, 4 Marek Proszek, 5 Katarzyna Faszczewska, 6 Bartosz Głogowski, 7 Piotr Kopa.

Quality

10 conversational test sentences per reader

voice	n	WER	CER	UTMOS	F0 spread [st]	chars/s	silence %
Jan Staszczyk	10	0.070	0.025	3.13	3.43	10.9	9
Radosław Krzyżowski	10	0.093	0.036	2.91	2.19	12.0	12
Wojciech Masiak	10	0.221	0.122	2.79	2.32	13.0	10
Bartosz Bielenia	10	0.105	0.040	3.02	1.75	11.4	13
Marek Proszek	10	0.093	0.050	2.91	2.69	12.2	8
Katarzyna Faszczewska	10	0.116	0.084	2.84	2.38	11.9	10
Bartosz Głogowski	10	0.093	0.084	3.32	2.95	12.7	9
Piotr Kopa	10	0.151	0.101	2.72	2.48	10.2	6

Training procedure (flattened)

Data: the Stage-2 target set of the Matcha recipe (8 Wolne Lektury readers, prose only, consistency-ranked, text/audio mismatch clips removed; piper_metadata.csv = file|reader|text). Base: pl_PL-darkman-medium from rhasspy/piper-checkpoints (single speaker) warm-started non-strictly into an 8-speaker model (--model.warmstart_ckpt). 15 000 steps, batch 32, fp32, validation 2 %, one H100 (~55 min). Export with piper.train.export_onnx (legacy exporter). Runtime: piper-tts (needs espeak-ng data; on macOS set ESPEAK_DATA_PATH=/opt/homebrew/share/espeak-ng-data).

Licence and attribution

• Weights: CC BY-SA 4.0 — trained on Wolne Lektury recordings (CC BY-SA 3.0 PL); every book, reader and director is listed in ATTRIBUTION.md. Base voice pl_PL-darkman-medium: data CC0 (OHF-Voice datasets), itself fine-tuned from the English lessac voice (public domain).
• Training code piper1-gpl and the piper-tts runtime are GPL-3.0; the ONNX weights and this card are not GPL. Using the voice through the Piper runtime means complying with GPL-3.0 for that software.
• Voices are personal attributes; the same caveats as for the Matcha release apply. Disclose synthetic speech.