Giving Voice to the Constitution: Low-Resource Text-to-Speech for Quechua and Spanish Using a Bilingual Legal Corpus

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Speech Technology · Depth: Expert, long

Summary

A unified pipeline has been developed to synthesize high-quality Quechua and Spanish speech for the Peruvian Constitution, utilizing three state-of-the-art text-to-speech (TTS) architectures: XTTS v2, F5-TTS, and DiFlow-TTS. The models were trained on independent Spanish and Quechua speech datasets of varying sizes and recording conditions, leveraging bilingual and multilingual TTS capabilities to enhance synthesis quality in both languages. This framework addresses data scarcity in Quechua through cross-lingual transfer while maintaining naturalness in Spanish. The project releases trained checkpoints, inference code, and synthesized audio for each constitutional article, providing a reusable resource for speech technologies in indigenous and multilingual contexts. This initiative aims to develop inclusive TTS systems for political and legal content in low-resource settings.

Key takeaway

For research scientists developing speech technologies for indigenous languages, this work demonstrates that high-quality, intelligible speech can be generated for low-resource languages like Quechua by leveraging cross-lingual transfer from high-resource languages such as Spanish. You should prioritize architectural design over model scale and consider DiFlow-TTS for its superior performance in balancing model size and synthesis quality, especially when data scarcity is a primary concern.

Key insights

Cross-lingual transfer in TTS effectively mitigates data scarcity for low-resource languages like Quechua.

Principles

• Cross-lingual learning outperforms model scaling in low-resource TTS.
• Architectural design is critical for efficient prosodic transfer.

Method

The method involves training XTTS v2, F5-TTS, and DiFlow-TTS on curated Quechua (40 hours) and Spanish (218 hours) corpora, applying duration-based filtering and morphological normalization for Quechua, and evaluating with UTMOS, SIM-O, WER, RMSEF0, and RMSEE.

In practice

• Use DiFlow-TTS for optimal quality in low-resource TTS.
• Employ bilingual training for data-scarce languages.

Topics

• Low-Resource Text-to-Speech
• Quechua Language
• Spanish Language
• Peruvian Constitution
• Cross-lingual Transfer

Code references

• coqui-ai/TTS
• SWivid/F5-TTS
• Tobertz-max/DiFlow-TTS

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Student

Related on AIssential

• Giving Voice to the Constitution: Low-Resource Text-to-Speech for Quechua and Spanish Using a Bilingual Legal Corpus — Cross-lingual transfer mitigates data scarcity in low-resource languages for high-quality TTS.
• Māori Text-to-Speech Model Spurns Big Tech’s Values — Indigenous communities can develop high-fidelity AI language models prioritizing data sovereignty and community ownership, offering an alternative to big tech.
• Low-resource Language Discrimination Towards Chinese Dialects with Transfer learning and Data Augmentation — Transfer learning and data augmentation effectively overcome scarce annotation resources for low-resource Chinese dialect discrimination.
• Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan — Synthetic data effectively bootstraps NMT structural learning for low-resource languages, but requires authentic data for semantic depth.
• KIT's Submission to Cross-Lingual Voice Cloning in IWSLT 2026 — Explicit language tags and lexical matching significantly improve cross-lingual voice cloning performance and reduce accent leakage.
• Generating Concept Lexicalizations via Dictionary-Based Cross-Lingual Sense Projection — A dictionary-augmented projection method improves cross-lingual WordNet expansion precision with minimal resources.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.