BSSD: Basque Synthetic Speech Dataset

364,870 utterances, 470.8 hours of synthetic Basque (euskara) read speech, 24 kHz mono FLAC, with normalised transcripts, voice metadata and per-utterance quality scores, in train / dev / test splits with held-out voices and sentences. Version 1.0-rc, 2026-10-17. Built by itzune as training data for Basque text-to-speech research; also usable for ASR data augmentation.

utterances / hours	364,870 / 470.8 h (mean 4.65 s)
voice identities	~18,400 ids, of which 180 have ≥ 1 h; 134 stable synthetic bank voices; 54 real speakers; 8 named default voices
intelligibility (ASR round trip)	CER median 0.0, p90 0.013, p99 0.027 on both Basque ASRs
naturalness proxy	UTMOS p10 3.16 / p50 3.51 / p90 3.90
text	236 h Wikipedia, 130 h Common Voice sentences, 75 h news (EusCrawl, CC BY-SA sites), 24 h generated numbers/dates, 4.5 h Wikisource, 0.5 h Tatoeba

How it was made

Teacher. All audio was generated with OmniVoice (k2-fsa/OmniVoice, commit c5fdb5cc, package built from git on 2026-10-15; 32 decoding steps, guidance 2.0, fp16). OmniVoice's code is Apache-2.0; its weights are CC-BY-NC (the authors cite training-data constraints such as Emilia) and it bundles Boson's Higgs Audio V2 tokenizer under the Boson community license. We state this provenance so that downstream users can make their own assessment; see "License" below.

Voices (voice_mode).

• clone_bank (213 h): 134 stable synthetic voices. OmniVoice generated 4,000 candidate voices in its auto and voice-design modes; candidates were kept only with CER ≤ 2 %, UTMOS in the top half, no clipping, then de-duplicated with WavLM x-vectors (cosine < 0.96, calibrated on real speakers) and balanced by pitch. Each kept clip is then used as the cloning prompt, so every bank voice is consistent across hundreds of sentences. Prompts are published in voices.jsonl.
• clone_named (77 h): eight named voices with a reserved share. Maider (16.4 h) and Antton (14.5 h) are the professional speakers of the HiTZ-Aholab studio corpus (CC BY 4.0), cloned from a 5–7 s studio clip; Nahia, Ane, Mikel, Jon, Unai, Aitor (7–8 h each) are synthetic voices chosen by listening among 84 voice-design candidates. The named voices were given extra hours on purpose so that models trained on the data reproduce them reliably.
• clone_real (126 h): clones of 54 real speakers who recorded for TTS: the 52 volunteers of OpenSLR 76 (CC BY-SA 4.0) and the two HiTZ-Aholab speakers. No Common Voice contributor was cloned.
• design (33 h) and auto (22 h): one-off voices from OmniVoice's attribute-controlled and random voice modes.

Text (text_src). Sentences were filtered (4–30 words, Basque alphabet, a vocabulary check against 380k trusted transcripts for web text, no abbreviations or foreign-letter tokens, no duplicates) and normalised by a rule-based Basque front-end (text_norm: cardinals and ordinals following Euskaltzaindia, case suffixes after digits such as 1990ean → mila bederatziehun eta laurogeita hamarrean, percentages, decimals, times, units, acronyms). Sentences with constructs the normaliser refuses to guess (ranges, numeric dates, fractions) were dropped. Sources: Basque Wikipedia (CC BY-SA), the Mozilla Common Voice Basque sentence pool (CC0), news and local media from the EusCrawl free release restricted to documents the publishers licensed CC BY-SA / CC BY (Berria, Goiena, Argia, Ataria, Guaixe, …), Basque Wikisource (CC BY-SA), Tatoeba (CC BY) and template-generated number/date sentences. Basque Parliament transcripts were excluded (lower-cased ASR fragments without punctuation).

Filtering (status = keep for every published row). Each clip was transcribed with xezpeleta/whisper-large-v3-eu (CT2) and HiTZ/stt_eu_conformer_transducer_large_v2; rows were kept with CER ≤ 3 % on both, hypothesis length ratio 0.7–1.3, 8–22 characters per second, UTMOS ≥ 3.0, speaker similarity to the prompt ≥ 0.90 (cloned rows), no clipping, no internal silence > 1.5 s, duration 1.5–20 s. About 85 % of generated rows passed; a per-voice cap of 1.5 h was applied to non-named voices; two rows were removed after listening (alegia mispronounced, one garbled utterance). Generated rows that failed are not included.

Does OmniVoice reproduce real Common Voice speakers? Common Voice (CC0) is in OmniVoice's training data. We compared every auto and design voice (4,000 clips) with one clip from each of the 9,678 Basque Common Voice v20 contributors using ECAPA-TDNN speaker embeddings. The nearest contributor to any random voice scored at most 0.78 (auto) while the 99th percentile of nearest different real contributors is 0.805; none of the 2,400 auto voices and 1 of 1,600 design voices (0.06 %, below chance) exceeded that line. The random voices are statistically new speakers.

Splits

split	utterances	hours	contents
train	347,559	448.0	everything else
dev	8,205	10.6	3 held-out bank voices (bank0020, bank0113, bank0026), 2 held-out real speakers (euf_07505, eum_08784), 0.5 % of one-off auto/design voices, 60 held-out sentences per named voice
test	9,106	12.1	3 held-out bank voices (bank0027, bank0083, bank0070), 2 held-out real speakers (euf_05679, eum_08886), 0.5 % of one-off auto/design voices, 60 held-out sentences per named voice

Every sentence occurs once in the dataset, so dev and test text is never seen in train. Held-out voices never occur in train; the named voices do (their dev/test rows test unseen text on a seen voice, the usual TTS preset condition). The plan is deterministic (splits.json, seed 2026).

Columns

column	meaning
id	eu_syn<run>_<shard>_<n>
audio	24 kHz mono FLAC, trimmed, peak-normalised
text / text_norm	sentence as written / as sent to the teacher (verbalised numbers etc.)
text_src	euwiki, cv, euscrawl, gen_numbers, wikisource, tatoeba
voice_mode, voice_id, prompt_src, instruct	see above; prompt_src is bank:*, named:*, oslr76:*, hitz:*
duration	seconds
cer_whisper, cer_conformer	ASR round-trip character error rate vs text_norm
utmos, spk_sim	predicted MOS; WavLM cosine similarity to the prompt (cloned rows)
teacher, num_step, guidance, status	generation parameters; keep

Extra files: voices.jsonl (bank voices and prompts), splits.json (split sizes and the held-out voice plan), manifest-{train,dev,test}-{a,b}.jsonl (file name, duration, transcript; the data was produced in two batches, a and b, hence two parquet series and two manifests per split).

Known limitations

• Synthetic read speech: prosody and timbre variety are those of the teacher; no spontaneous speech.
• Occasional Spanish-style readings of Basque orthography survive ASR filtering (e.g. ge/gi read as /xe, xi/ in ~0.1 % of such rows; the name Julen read with a fricative). Numbers are verbalised by rules and may differ from a speaker's choice in rare cases.
• Mean duration is short (4.65 s); long-form prosody is under-represented.
• The non-named voices come from a model whose Basque training audio is mostly crowd-sourced, so their acoustic quality is "clean crowd microphone" rather than studio, except for Maider, Antton and the real-speaker clones.

License and attribution

The dataset is released under CC BY-SA 4.0. Please credit: BSSD (Basque Synthetic Speech Dataset), itzune, 2026, and the sources it builds on: k2-fsa OmniVoice (Zhu et al., 2026, arXiv:2604.00688); Wikimedia (Basque Wikipedia and Wikisource); Mozilla Common Voice; the EusCrawl corpus (Artetxe et al., 2022) and the publishers Berria, Goiena, Argia and others; Google / OpenSLR 76 (Kjartansson et al., 2020); HiTZ Zentroa – Aholab (studio corpus, speakers Maider and Antton); Tatoeba. Users should be aware that the teacher model's weights are licensed CC-BY-NC by their authors; we document the full chain above rather than claim a cleaner one.

Citation

@misc{bssd2026,
  title  = {BSSD: Basque Synthetic Speech Dataset},
  author = {itzune},
  year   = {2026},
  url    = {https://huggingface.co/datasets/itzune/bssd}
}