tts-pl-matcha-nowl — Polish Matcha-TTS trained without Wolne Lektury

Companion release to hvsr-robotics/tts-pl-matcha-v2 for deployments that must avoid audiobook-actor voices. Same architecture (Matcha-TTS, 20.8 M parameters, 625 speaker rows, no style tokens), trained on Multilingual LibriSpeech Polish (LibriVox volunteers), YODAS Polish YouTube speech and AZON spontaneous speech. Quality is lower than the Wolne Lektury model (UTMOS ≈ 3.0–3.2 vs 3.3–3.5) because the source recordings are compressed 16 kHz audio.

Files

path	what
runs_wcss/matcha_nowl/checkpoints/final.ckpt	Lightning checkpoint (epoch 19, 13.8k steps)
export/matcha_nowl_t2.onnx, export/matcha_nowl_t4.onnx	acoustic model + HiFi-GAN, 2 / 4 ODE steps; inputs x, x_lengths, scales, spks
checkpoints/g_02500000	HiFi-GAN universal vocoder
data_v2/mix_nowl/speaker_map.json	speaker id → source / speaker (0–10 MLS, 11–603 YODAS channels, 604–624 AZON)
samples/, ATTRIBUTION.md, LICENSE, RECIPE.md	as in tts-pl-matcha-v2; UI and tools: tts-pl-playground

Recommended voices: speaker 0 (MLS narrator, UTMOS 3.11) or the blend 0.5*0 + 0.3*11 + 0.2*604 (UTMOS 3.16). Most YODAS speaker rows have very few training clips; treat ids 11–603 as experimental.

Quality

10 conversational test sentences (spk0 = MLS narrator, spk11 = YODAS, spk604 = AZON, mix0 = 0.5·0 + 0.3·11 + 0.2·604)

voice	n	WER	CER	UTMOS	F0 spread [st]	chars/s	silence %
spk0	10	0.058	0.025	3.11	1.65	11.3	11
spk11	10	0.058	0.067	2.98	1.97	13.4	6
spk604	10	0.070	0.076	2.99	1.28	13.7	5
mix0	10	0.058	0.071	3.16	2.27	12.3	11

Training procedure (flattened)

1. MLS Polish: 11 speakers, 8 h cap per speaker (25 h), re-transcribed with Whisper large-v3 (punctuation and casing transferred onto the original words; fixes dropped "j" in one speaker). YODAS pl000: DNSMOS filter, two-model Whisper transcript agreement. AZON spontaneous speech as is. All 22.05 kHz mono, peak-normalised.
2. Pre-train on MLS from the VCTK warm start (weights only): 60k steps, then 10k steps on the Whisper-fixed text.
3. Fine-tune on MLS + YODAS + AZON (625 speakers, 44k lines, source-balanced): 13.8k steps, Adam 1e-4, batch 64, bf16.
4. Export as above. Same cleaner, symbol table and evaluation protocol as the v2 model (see RECIPE.md, sections 1, 5, 6).

Licence and attribution

• Weights: CC BY-SA 4.0 (AZON is CC BY-SA 4.0; MLS CC BY 4.0; YODAS CC BY 3.0 — attribution to the 894 YouTube channels is listed in ATTRIBUTION.md). Warm start matcha_vctk.ckpt (MIT; VCTK CC BY 4.0); HiFi-GAN universal (MIT).
• YODAS licences rest on the uploaders' CC BY declarations as collected by the YODAS authors; verify before commercial use.
• Voices are personal attributes; prefer blended voices with neutral names. Disclose synthetic speech to listeners.