machinekind/polish-synthetic-speech-voxcpm2

Polish synthetic speech — VoxCPM2, 4 voices

Newer version: polish-synthetic-speech-voxcpm2-v2.

31867 clips, 59.6 h of modern, diverse Polish: the same texts read by 4 synthetic voices, generated with VoxCPM2 (OpenBMB, Apache-2.0) and quality-gated clip by clip. Made to distil a large TTS into small on-device ones for a humanoid robot (Matcha-TTS-PL, try it in the browser).

voice	gender	clips	hours	mean UTMOS	similarity to its anchor	designed from the description
male_voxcpm_1	male	7979	15.1	3.64	0.923	Male, teenager, energetic, optimistic, ready to help, speaks crip Polish.
female_voxcpm_1	female	7960	15.1	3.77	0.919	Female, teenager, energetic, optimistic, confident, ready to help, speaks crip Polish, mid/high-pitch timbre, mid/fast tempo.
male_voxcpm_2	male	8015	14.8	3.74	0.911	Voice of a dog, fast, energetic, crazy, mid/high pitch.
female_voxcpm_2	female	7913	14.6	3.64	0.947	Female, teenager, energetic, optimistic, confident, ready to help, speaks crip Polish, mid/high-pitch timbre, mid/fast tempo.

• Synthetic voices, not people. Each voice was designed from a text description, frozen as a VoxCPM2 prompt cache (voices/) and used for every clip of that voice.
• Parallel. All voices read the same texts (text_id is shared), and the train/validation split is the same for every voice. A voice lacks a text only when all 3 of its takes failed the quality gate (≈ 2 %).
• Both grammatical genders, independent of the voice. Polish marks the speaker's gender in first-person forms (zrobiłem / zrobiłam). Of the texts with such forms, a random half was rewritten with feminine forms and the rest keep masculine ones; every voice reads both. Column speaker_form: 1870 masculine, 1762 feminine, 28235 neutral clips.
• What a humanoid assistant says. Speech written for this dataset in 18 everyday domains (home, care, kids, office, travel, emergencies, small talk, stories…), modern corpora (NKJP, Wikinews, KPWr, ParlaMint, OpenAssistant, Wikibooks, Wikivoyage), reviewed YouTube transcripts and template lines with numbers, times and dates. No archaic literature.
• Questions of every kind (10973 clips end with "?"): wh-questions with every question word at different distances from the "?", after lead-ins ("Zapytacie, po co…"), yes/no questions with and without czy, tag and alternative questions, and questions closing a passage.
• Long utterances. 36.7 h are clips of 10 s or more: passages of 2–6 sentences, for prosody across sentences.
• English words inside Polish (meeting, feedback, YouTube…), read the English way; text_marked puts them in braces (Po {meeting}u wyślę ci {feedback}.) for a mixed Polish/English phonemizer.
• Clean text. Numbers, dates and abbreviations written out as words; source sentences cut mid-thought were completed (cut sentences teach a TTS wrong sentence endings).

Contents

category	clips	hours
assistant speech (passage)	7690	31.93
assistant speech (sentence)	7582	9.13
spoken (YouTube transcripts)	5075	5.18
assistant templates	3164	2.32
modern corpora (nkjp)	492	2.15
assistant speech (short)	4736	2.03
modern corpora (kpwr)	295	1.36
modern corpora (wikinews)	314	1.34
modern corpora (parlamint)	279	1.23
Polish with English words	1804	1.05
modern corpora (oasst2)	219	0.90
modern corpora (wikibooks)	115	0.52
modern corpora (wikivoyage)	102	0.42

split	clips	hours
train	31226	58.37
validation	641	1.21

Audio: FLAC, 48000 Hz mono, 16-bit, leading/trailing silence trimmed (80 ms kept). Columns: file_name, text_id (shared by all voices), voice, voice_gender, speaker_form (masculine / feminine / neutral first-person forms in the text), text, text_marked (English spans in braces), category, duration, style (pitch-range tercile × tempo tercile within the voice, see style_map.json; the question bit is always 0), utmos, speaker_similarity (to the voice's anchor), source, license, url, author, title, modified.

from datasets import load_dataset
ds = load_dataset("machinekind/polish-synthetic-speech-voxcpm2", "female_voxcpm_1", split="train")   # one voice
ds = load_dataset("machinekind/polish-synthetic-speech-voxcpm2", split="train")                      # all voices (filter by "voice")
print(ds[0]["voice"], ds[0]["text"], ds[0]["audio"]["sampling_rate"])

How it was made

1. Texts — see the categories above; every corpus or transcript sentence was reviewed by Claude (keep / minimally complete or correct / drop: cut sentences, garbled transcripts, preaching, partisan politics, crime and anything unsuitable for a household robot were removed).
2. Speech — VoxCPM2 with each frozen voice (anchor + continuation conditioning, 10 diffusion steps, cfg 2.0), one seed per sentence and take.
3. Quality gate, per clip — Whisper large-v3-turbo transcript vs text (letter CER ≤ 8 %, length ratio 0.8–1.25, numbers suppressed to words), UTMOS ≥ 3.2, speaker similarity to the anchor ≥ 0.80 (≥ 0.65 below 2.5 s), tempo, and median pitch within 0.6–1.75 × the voice's own anchor pitch. Up to 3 takes per sentence and voice; about 89 % of takes passed. A take whose English words came back intact was preferred.
4. Text = exactly what VoxCPM2 read. Punctuation is never changed after generation: VoxCPM2 decides the intonation (wh-questions mostly fall, yes/no questions mostly rise), and the text keeps the "?" it read.

Known artefacts: occasional teacher mispronunciations that Whisper accepts, a few Polish-pronounced loanwords marked as English, rare unnatural pauses in long passages.

Licence and attribution

CC BY-SA 4.0. Text sources and their licences (per text in credits.jsonl and the metadata):

source	texts	licences
written for this dataset with Claude	5072	CC BY-SA 4.0 (original text)
YODAS pl000	1297	CC BY 3.0
written for this dataset	1250	CC BY-SA 4.0 (original text)
NKJP 1M subcorpus 1.2	131	CC BY
Wikinews PL	87	CC BY 2.5 (CC BY 4.0 for later revisions)
KPWr 1.3	80	CC BY, CC BY 2.0 PL, CC BY 2.5 ES, CC BY 2.5 PL, CC BY 3.0 PL, CC BY-SA 2.5 PL, CC BY-SA 3.0, CC BY-SA 3.0 PL, CC Copyright-Only Dedication* (based on United States law) or Public Domain Certification (retired)
ParlaMint-PL 5.0	74	CC BY 4.0
OpenAssistant oasst2	58	Apache-2.0
Wikibooks PL	33	CC BY-SA 4.0
Wikivoyage PL	29	CC BY-SA 4.0

ATTRIBUTION.md lists every source document, author or speaker and link, and explains the licence compatibility. OpenAssistant texts and the VoxCPM2 model are Apache-2.0 (LICENSES/Apache-2.0.txt). The voices are synthetic: label speech made with them as synthetic and do not present them as real people.

Citation

@misc{machinekind_polish_synthetic_speech_2026,
  title  = {Polish synthetic speech (VoxCPM2, multiple synthetic voices)},
  author = {Machinekind},
  year   = {2026},
  url    = {https://huggingface.co/datasets/machinekind/polish-synthetic-speech-voxcpm2}
}

Please also cite the corpora listed in ATTRIBUTION.md (NKJP, ParlaMint, KPWr, OpenAssistant, YODAS) and VoxCPM2.