PoliticalNLP 2026 · 2026 · Workshop

John E. Ortega et al. (Northeastern University / Universitat Pompeu Fabra / Barcelona Supercomputing Center) · → Paper · Demo: ? · Code: ?

Trains three existing open TTS systems (XTTS-v2, F5-TTS, DiFlow-TTS) on pooled Spanish and Southern Quechua speech to voice the Peruvian Constitution in Quechua, and compares them with automatic metrics only; the smallest model, DiFlow-TTS, scores best on UTMOS, WER and prosody error, while F5-TTS has the highest speaker similarity.

Problem

Quechua, Peru’s largest indigenous language, has an official translation of the national constitution but little speech technology to make that text audible for radio, legal aid, or screen readers. Transcribed Quechua speech is scarce, noisy, and orthographically variable, so a Quechua TTS system trained on its own data alone is hard to build. The authors also want a domain-matched synthetic audio resource that could later feed Quechua ASR and speech translation work, building on earlier findings that TTS-generated data helps Quechua ASR (§2).

Method

The work is a data and training pipeline around three off-the-shelf architectures rather than a new model (§3.2).

Data (§3.1). Quechua training data combine the Siminchik and Lurin corpora (about 97.5 and 83.3 hours of Southern Quechua, roughly 180 hours in total). Transcriptions are normalized with a morphological parser into standardized Southern Quechua forms, misaligned samples are removed, and all segments shorter than one second plus low-quality clips (music, noise, silence) are discarded. This filtering removes about 140 hours, leaving about 40 hours of Quechua. Spanish data come from crowdsourced and professionally recorded multi-dialect Latin American and Peninsular Spanish TTS corpora plus TEDx Spanish, about 218 hours after resampling, amplitude normalization, and sentence segmentation. Constitution text in both languages is segmented into article-aligned JSON, with numbers, dates, legal abbreviations, and Roman numerals expanded; Quechua uses a rule-based grapheme-to-phoneme front end.

Models (§3.2). All three systems are trained or fine-tuned on the bilingual data following their official recipes on H100 GPUs:

  • XTTS-v2 (470M), a GPT-style autoregressive model, initialized from pretrained checkpoints with standard Coqui hyperparameters, AdamW, cosine schedule, and FP16.
  • F5-TTS (336M), a flow-matching model with a ConvNeXt text encoder and Diffusion Transformer backbone, keeping its default objective and Sway Sampling inference.
  • DiFlow-TTS (164M), which applies discrete flow matching over factorized codec tokens with separate prosody and acoustic prediction heads.

Learning rates of 1e-4 and 5e-5, weight decay, gradient clipping, and batch sizes are tuned per model, with early stopping on validation loss and audio quality. Which model receives which setting, the number of steps, and whether each model is fine-tuned or trained from scratch are not reported per system. Output is 22.05 to 24 kHz audio.

Evaluation (§3.3). Evaluation is automatic only and covers Quechua audio synthesized from the constitution text: UTMOS as a naturalness proxy, SIM-O as speaker-embedding cosine similarity across utterances, WER from the QUESPA self-supervised Quechua ASR model, and RMSE of F0 and energy against reference signals.

Key Results

Table 1 (verified against table-1.png) reports one row per system, all trained bilingually:

ModelParamsUTMOS ↑SIM-O ↑WER ↓RMSE F0 ↓RMSE E ↓
XTTS-v2470M3.220.530.1921.030.021
F5-TTS336M3.230.600.1915.170.017
DiFlow-TTS164M3.310.490.1610.240.011

DiFlow-TTS leads on four of five quality metrics (UTMOS, WER, and both prosody errors) despite being the smallest model, but it has the lowest speaker similarity of the three (0.49 against 0.60 for F5-TTS and 0.53 for XTTS-v2). The UTMOS spread across systems is narrow (3.22 to 3.31), and XTTS-v2 and F5-TTS tie on WER. The prosody errors differ more clearly, roughly halving from XTTS-v2 to DiFlow-TTS.

The authors conclude that cross-lingual learning matters more than model scale and that Spanish data “enables” intelligible Quechua synthesis (§4). Table 1 contains no Quechua-only baseline, no ground-truth or resynthesis reference row, and no Spanish results, so neither conclusion is directly tested: the comparison is only among three bilingual systems that differ in architecture, size, and training recipe all at once.

Novelty Assessment

There is no architectural novelty; the three models are used as published. The contribution is applied: a bilingual Quechua and Spanish training setup with Quechua-specific normalization and filtering, a legal-domain text front end, and an automatic comparison of three open architectures for a low-resource indigenous language. The promised release of checkpoints, inference code, and constitution audio is a potential resource, but §7 states the data and prompts were omitted at submission time, and no URL is given in the paper.

Field Significance

low — The paper provides a small data point on adapting current open TTS systems (autoregressive, continuous flow matching, discrete flow matching) to Southern Quechua through pooled training with Spanish, showing that all three reach comparable automatic naturalness while differing mainly in prosody error and speaker similarity. Without a monolingual control or human evaluation, it serves as an engineering report and a pointer to a possible Quechua legal-domain speech resource rather than as evidence about cross-lingual transfer itself.

Claims

  • supports: When adapting TTS to a low-resource language, smaller non-autoregressive flow-matching models can match or exceed a larger autoregressive model on automatic naturalness, intelligibility, and pitch accuracy.

    Evidence: Under the same bilingual Spanish and Quechua data, DiFlow-TTS (164M) obtains UTMOS 3.31, WER 0.16, and F0 RMSE 10.24, against 3.22, 0.19, and 21.03 for XTTS-v2 (470M); F5-TTS (336M) sits between them on F0 RMSE (15.17). (§4, Table 1)

  • complicates: Rankings of low-resource TTS systems depend on which automatic metric is used, since speaker similarity can move opposite to naturalness and prosody accuracy.

    Evidence: DiFlow-TTS is best on UTMOS, WER, and both RMSE metrics but worst on SIM-O (0.49), while F5-TTS has the highest SIM-O (0.60) with middling UTMOS (3.23). (§4, Table 1)

  • supports: Found speech corpora for low-resource languages can lose most of their volume under TTS-oriented duration and quality filtering.

    Evidence: Removing segments under one second and noisy, musical, or silent clips cuts the combined Siminchik and Lurin Quechua data from about 180 hours to about 40 hours. (§3.1)

  • complicates: Claims that cross-lingual pooling with a high-resource language improves low-resource TTS need a target-language-only control, because architecture, size, and recipe differences confound comparisons among pooled systems.

    Evidence: The paper attributes Quechua quality to Spanish transfer and states that cross-lingual learning outweighs scale, but all reported systems are bilingual and no Quechua-only or reference row is given. (§4, Table 1)

Limitations and Open Questions

No control for the central claim

The paper’s headline conclusion, that cross-lingual transfer from Spanish rather than model scale drives Quechua quality, is not tested: there is no Quechua-only training baseline, no human listening test, and no ground-truth reference row, and the three compared systems differ in architecture, size, and recipe simultaneously.

  • Evaluation is entirely automatic. UTMOS is an English-trained MOS predictor whose validity for Quechua is not examined, and the WER depends on a single Quechua ASR model.
  • Only Quechua output is evaluated, although the system and stated goal are bilingual; Spanish quality is not reported.
  • Training details are given as ranges shared across models (two learning rates, “tens to hundreds of thousands of steps”), so per-model settings, which models are fine-tuned from pretrained checkpoints, and the vocoder or codec used for each system are not reported.
  • How the SIM-O reference is chosen and what the RMSE reference signals are for constitution text (which has no natural recordings) is not described.
  • The paper states that checkpoints and audio are released, but §7 says data and prompts were withheld for anonymity, and no repository link appears in the paper.
  • The dialect scope (Cusco Collao influence) and the suitability of synthesized legal text for formal or court settings are noted by the authors as open (§5).

Wiki Connections

  • Multilingual TTS — Trains three TTS systems jointly on Spanish and Southern Quechua to synthesize the low-resource language through cross-lingual pooling.
  • Flow Matching — Fine-tunes two flow-matching systems, continuous (F5-TTS) and discrete over codec tokens (DiFlow-TTS), which outperform the autoregressive baseline on prosody error.
  • Autoregressive Codec TTS — Adapts the autoregressive XTTS-v2 to Quechua as the largest compared system, which has the highest F0 and energy errors.
  • XTTS — One of the three adapted systems, initialized from pretrained checkpoints with the Coqui recipe.
  • F5-TTS — One of the three adapted systems, with default flow-matching objective and Sway Sampling; highest speaker similarity in Table 1.
  • DiFlow-TTS — One of the three adapted systems and the best performer on UTMOS, WER, and prosody error at the smallest size.
  • SeamlessM4T — Cited as a multilingual pretrained model behind gains in Quechua to Spanish speech translation, motivating the reuse of the synthesized audio for speech translation.