IWCLUL 2025 · 2025 · Workshop
Katri Hiovain-Asikainen et al. (UiT The Arctic University of Norway) · → Paper · Demo: ✓ · Code: ✗
Builds and releases the first text-to-speech system for South Sámi, a severely endangered language with 500-600 speakers, by restoring and training on roughly 10.5 hours of 1989-1993 archival recordings of a single deceased speaker.
Problem
South Sámi is spoken by an estimated 500-600 people across Norway and Sweden, and its written standard was only established in 1978, so many speakers are not fully comfortable reading the language even though they speak it. No TTS system existed for South Sámi before this work, and none of the standard ingredients for building one were available: there is no sizable clean speech corpus, no pre-trained South Sámi acoustic model to fine-tune, and transfer learning from related Sámi languages was judged unsuitable because South Sámi differs substantially from North and Lule Sámi in both linguistics and orthography. The only usable speech material was a set of degraded archival recordings, made between 1989 and 1993, of a single speaker who had since died. Established TTS evaluation protocols, designed for languages with large populations of available raters, also do not transfer cleanly to a community this small without risking research fatigue and undue burden on speakers.
Method
The team built a South Sámi speech corpus from archival recordings of Anna Jacobsen, a South Sámi language advocate recorded by the Norwegian national broadcaster NRK and on audiobook cassettes between 1989 and 1993, spanning news, biblical readings, fairy tales, and autobiographical storytelling. Written consent for open-source publication, TTS use, and royalty waiver was obtained from Jacobsen’s descendants, and the project followed FAIR and CARE principles for Indigenous data. Text was aligned to existing published transcriptions where available, or manually transcribed by three project members where not; roughly ten hours of audio required about one hundred hours of transcription effort. Sentence boundaries were found by force-aligning audio and text with a WebMAUS pipeline (using a Finnish acoustic model, since no Sámi model exists) and manually checked, yielding 10.5 hours across 4,670 sentences.
The audio was then denoised and enhanced with Resemble-Enhance, level-normalized, and resampled to 22.05 kHz. Rather than discarding the noisier portions, the authors split the data by manually assessed quality into a 2-hour “good quality” subset (speaker ID 1) and an 8.5-hour “lower quality” subset (speaker ID 0), and trained a FastPitch acoustic model in a multi-speaker configuration over both subsets, using speaker ID 1 for final synthesis. FastPitch, a non-autoregressive Transformer architecture with explicit duration and pitch prediction, was chosen instead of large zero-shot models such as VALL-E 2 because prior low-resource TTS work (e.g., on Lithuanian and Võro) showed that Transformer-based TTS can reach intelligible quality from as little as 1-2 hours of data, with little benefit from cross-lingual transfer. The model was trained for 830 epochs (14K steps) on the Saga supercomputer with an effective batch size of 256 and learning rate 0.1. For inference, the team used the UnivNet vocoder rather than HiFi-GAN, citing UnivNet’s multi-resolution spectrogram-loss training as giving better generalization across recording conditions and lower inference latency for real-time use.
Text normalization is handled separately from the acoustic model by the existing GiellaLT infrastructure, a rule-based, finite-state pipeline (built on HFST and VISLCG3) shared across all Divvun language tools. Raw text is tokenized, morphologically analyzed, disambiguated with constraint-grammar rules, and normalized (compound reconstruction, numeral lexicalization) before being passed to the TTS model. The resulting model was compiled to TorchScript and integrated as a system-level voice on macOS and Windows, usable as a screen-reader voice and, on macOS, within LibreOffice.
Key Results
The South Sámi voice was released on 30 October 2024. For initial evaluation, the authors held out 15 sentences from training, synthesized them, and paired them with ground-truth recordings of the same sentences in a 30-sample online survey using a 5-point Likert scale across three dimensions: Pronunciation & Rhythm, Pleasantness, and Clarity. Across 12 paired items analyzed with Wilcoxon signed-rank tests, synthesized speech showed no statistically significant difference from ground truth on Clarity (p ≈ 0.33) or Pronunciation & Rhythm (p ≈ 0.52), but a significant gap on Pleasantness (p ≈ 0.0039), with ground-truth recordings rated higher. A separate qualitative review by one native South Sámi speaker outside the author team flagged three specific issues: year numbers sometimes rendered digit-by-digit rather than as natural number words, vowel duration and speaking rate that could be improved, and more accurate pronunciation of hyphenated versus unhyphenated compounds. The authors report positive informal community reception, including interest in integrating the voice into flashcard-based language learning tools and its use as a teaching and elder-care aid, alongside an ongoing anonymous feedback survey launched roughly a year after release.
Novelty Assessment
The contribution is not architectural: FastPitch and UnivNet are both used off the shelf, and the text-processing front end is an existing shared infrastructure component. The genuine contributions are, first, a documented and ethically grounded pipeline for turning a small quantity of severely degraded, decades-old archival recordings from a single deceased speaker into a usable TTS training corpus, including audio restoration, alignment-assisted transcription, and quality-stratified multi-speaker training rather than discarding the lower-quality material; and second, being the first working TTS system for a critically endangered language with a community numbering in the low hundreds, built and evaluated under explicit descendant consent and CARE-aligned data governance. The reduced-scale evaluation protocol, deliberately scoped down from standard TTS listening-test practice to limit burden on a small speaker community, is presented as a considered methodological choice rather than a shortfall.
Field Significance
moderate - this paper demonstrates that an existing low-data TTS architecture (FastPitch) can be applied successfully to archival, single-speaker, quality-mixed recordings for a language with only a few hundred speakers, and it documents a reusable data-curation and consent process (informed descendant consent, quality-stratified corpus splitting, community-scaled evaluation) that other endangered-language TTS efforts could adapt. Its primary value is as a dataset-and-process contribution and an engineering integration of existing components for a previously unaddressed language, rather than a new modeling technique.
Claims
- supports: Archival recordings, even when severely degraded and drawn from a single speaker recorded decades earlier, can be systematically restored and combined with quality-stratified multi-speaker training to produce an intelligible neural TTS voice for a language with no prior training corpus.
Evidence: 10.5 hours of 1989-1993 recordings, denoised with Resemble-Enhance and split into a 2-hour “good” and 8.5-hour “lower quality” subset trained as separate speaker IDs in FastPitch, produced a South Sámi voice with no statistically significant Clarity or Pronunciation & Rhythm gap versus ground-truth recordings (p ≈ 0.33 and p ≈ 0.52). (§3.2, §3.3, §4.2)
- complicates: A TTS voice trained on quality-mixed archival material can match ground-truth recordings on intelligibility measures while still lagging on perceptual pleasantness.
Evidence: Wilcoxon signed-rank tests over 12 paired evaluation items found a statistically significant difference favoring ground truth only for the Pleasantness dimension (p ≈ 0.0039), with no significant gap for Clarity or Pronunciation & Rhythm. (§4.2)
- supports: Decoupling text normalization into a shared, rule-based finite-state pipeline lets improvements to text processing propagate to synthesis quality without retraining the acoustic model.
Evidence: The GiellaLT/HFST+VISLCG3 pipeline used for tokenization, disambiguation, and numeral/abbreviation normalization is shared across all Divvun language tools; the authors note that fixing text-processing issues such as numeral rendering, identified in the qualitative review, would improve output without touching the FastPitch model itself. (§3.4, §4.4)
- complicates: Standard large-panel TTS subjective-evaluation protocols are difficult to apply directly to critically endangered languages, because the available rater pool is a small fraction of an already tiny speaker population.
Evidence: The authors explicitly depart from evaluation norms designed for majority languages, citing the 500-600-speaker South Sámi population and risk of research fatigue as reasons to use a reduced 15-sentence test set and a single external native-speaker qualitative reviewer rather than a large-panel listening study. (§4.2)
Limitations and Open Questions
The reported evaluation rests on a small held-out set (15 sentences, 12 items entering the paired statistical test) and a single external native-speaker qualitative reviewer. The authors themselves argue that standard large-panel TTS evaluation cannot be ethically or practically applied given the size of the South Sámi speaker community, so these results should be read as preliminary community-scoped evidence rather than a statistically powered benchmark.
The training corpus is drawn from a single, now-deceased speaker, so the released voice cannot be validated or compared against alternative South Sámi speakers, and there is no existing South Sámi TTS baseline to compare against. Transfer learning from related Sámi languages was deliberately not attempted, so the system offers no multilingual or cross-lingual capability by design; the authors describe extending to Swedish/Norwegian loanword handling and additional voices as future work rather than something demonstrated here. The modified FastPitch training code was, at the time of writing, not yet publicly released, pending documentation. The qualitative review also surfaced concrete, unresolved text-processing issues (numeral rendering, compound hyphenation, vowel duration) that the authors flag for future iteration rather than having fixed in the current release.
Wiki Connections
- GAN Vocoder — the released voice uses UnivNet, a GAN-based vocoder with multi-resolution spectrogram discriminators, chosen over HiFi-GAN for better generalization across the archival material’s varied recording conditions and for low-latency inference.
- Transformer Encoder-Decoder TTS — the acoustic model is FastPitch, a non-autoregressive Transformer TTS architecture selected specifically for its ability to reach intelligible quality from only a few hours of speech.
- Evaluation Metrics — the paper argues that standard TTS evaluation protocols, built for high-resource languages with large rater pools, do not transfer directly to a severely endangered language and proposes a reduced-scale, community-conscious alternative.
- Subjective Evaluation — the voice is assessed through a genuine listening test, with 5-point Likert ratings compared against ground truth via Wilcoxon signed-rank tests and a native-speaker qualitative review.
- VALL-E 2 — cited as representative of the massive-dataset, zero-shot TTS paradigm that the authors explicitly reject as impractical given only about 10 hours of usable archival audio for South Sámi.
- Emilia — cited alongside VALL-E 2 as representative of the large multilingual-corpus trend in TTS that motivates this paper’s low-resource, single-language alternative.
- HiFi-GAN — used as the comparison point when justifying the choice of the UnivNet vocoder for the South Sámi voice.