MayaVoice is a machine translation system between Spanish and 14 Mayan languages of Guatemala. It is the first open-source large language model fine-tuned specifically for this linguistic family, designed to bridge the communication gap between Spanish-speaking services and the 6+ million Maya speakers across Guatemala.

Motivation

Guatemala is a multilingual country where 14 Mayan languages hold official recognition. Yet the linguistic barrier between Spanish and these languages severely limits access to healthcare, education, justice, and government services for millions of Maya speakers. MayaVoice aims to reduce this gap through AI-powered translation.

Supported Languages and Evaluation

The model was evaluated on a held-out test set of 300 sentences unseen during training:

Language	Family	Speakers (est.)	N	BLEU	chrF
Mam	Mamean	530,000	70	47.07	61.30
Kaqchikel	K'ichean	450,000	57	41.34	56.29
Chuj	Q'anjob'alan	65,000	11	31.32	54.78
Q'eqchi'	K'ichean	800,000	22	27.25	52.03
Q'anjob'al	Q'anjob'alan	170,000	23	23.38	43.92
K'iche'	K'ichean	1,100,000	18	21.32	38.57
Tektiteko	Mamean	5,000	20	15.76	43.78
Sipakapense	K'ichean	8,000	7	15.75	48.87
Awakateko	Mamean	20,000	12	15.46	34.46
Poqomchi'	K'ichean	115,000	7	15.15	31.88
Poqomam	K'ichean	50,000	24	11.21	48.97
Achi	K'ichean	150,000	4	8.30	44.68
Itza'	Yucatecan	1,000	7	4.02	17.21
Tz'utujil	K'ichean	90,000	18	2.86	28.81
Weighted avg		~3.5M	300	40.28	55.07

Note on metrics: BLEU measures exact n-gram overlap; chrF measures character-level similarity, which is more appropriate for agglutinative languages like those of the Mayan family. Languages with lower N have wider confidence intervals.

Linguistic Observations

Mamean languages (Mam, Tektiteko, Awakateko) and K'ichean languages (Kaqchikel, K'iche', Q'eqchi') show the strongest results, correlating with greater availability of parallel training data.
Itza' (Yucatecan family) has the lowest performance, consistent with its status as a critically endangered language with very few digitized texts.
chrF scores are substantially higher than BLEU across all languages, which is expected for morphologically complex languages where partial word matches capture translation quality more faithfully.
Performance variation across the K'ichean branch (Kaqchikel 41.34 BLEU vs. Tz'utujil 2.86) highlights that genetic relatedness alone does not predict translation quality — data availability is the dominant factor.

Demo

Try MayaVoice live: 🌽 MayaVoice Space

Limitations

Domain bias: The training corpus has a non-uniform domain distribution, which may affect quality on everyday conversational language.
Low-resource languages: Itza', Tz'utujil, and Achi have significantly fewer training resources, reflected in their metrics.
Hallucination risk: Like all generative models, MayaVoice may produce plausible-looking but incorrect translations. Human verification is recommended for critical use cases.
Limited test set: 300 test sentences is a small evaluation set; per-language metrics with N < 10 are indicative, not conclusive.

Ethical Use

MayaVoice was developed to preserve and facilitate access to the Mayan languages of Guatemala. It should not be used to replace human translators in contexts where accuracy is critical (legal, medical), but rather as a support and accessibility tool.

Citation

bibtex

@misc{regalado2025mayavoice,
  author = {Regalado Cardoso, Daniel},
  title = {MayaVoice: Machine Translation for 14 Mayan Languages of Guatemala},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/DanielRegaladoCardoso/mayavoice-llama3.1-8b-lora-v2}
}

Contact

Developer: Daniel Regalado Cardoso
Institution: University of Miami
Demo: MayaVoice Space