Saved in:
Bibliographic Details
Main Authors: Chary, Luis Felipe, Ramirez, Miguel Arjona
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.05863
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915483173781504
author Chary, Luis Felipe
Ramirez, Miguel Arjona
author_facet Chary, Luis Felipe
Ramirez, Miguel Arjona
contents We present LatinX, a multilingual text-to-speech (TTS) model for cascaded speech-to-speech translation that preserves the source speaker's identity across languages. LatinX is a 12-layer decoder-only Transformer trained in three stages: (i) pre-training for text-to-audio mapping, (ii) supervised fine-tuning for zero-shot voice cloning, and (iii) alignment with Direct Preference Optimization (DPO) using automatically labeled pairs based on Word Error Rate (WER) and speaker-similarity metrics. Trained on English and Romance languages with emphasis on Portuguese, LatinX with DPO consistently reduces WER and improves objective similarity over the fine-tuned baseline. Human evaluations further indicate stronger perceived speaker similarity than a strong baseline (XTTSv2), revealing gaps between objective and subjective measures. We provide cross-lingual analyses and discuss balanced preference signals and lower-latency architectures as future work.
format Preprint
id arxiv_https___arxiv_org_abs_2509_05863
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle LatinX: Aligning a Multilingual TTS Model with Direct Preference Optimization
Chary, Luis Felipe
Ramirez, Miguel Arjona
Computation and Language
We present LatinX, a multilingual text-to-speech (TTS) model for cascaded speech-to-speech translation that preserves the source speaker's identity across languages. LatinX is a 12-layer decoder-only Transformer trained in three stages: (i) pre-training for text-to-audio mapping, (ii) supervised fine-tuning for zero-shot voice cloning, and (iii) alignment with Direct Preference Optimization (DPO) using automatically labeled pairs based on Word Error Rate (WER) and speaker-similarity metrics. Trained on English and Romance languages with emphasis on Portuguese, LatinX with DPO consistently reduces WER and improves objective similarity over the fine-tuned baseline. Human evaluations further indicate stronger perceived speaker similarity than a strong baseline (XTTSv2), revealing gaps between objective and subjective measures. We provide cross-lingual analyses and discuss balanced preference signals and lower-latency architectures as future work.
title LatinX: Aligning a Multilingual TTS Model with Direct Preference Optimization
topic Computation and Language
url https://arxiv.org/abs/2509.05863