PiDA: Phonetically-Informed Data Augmentation for Robust Vietnamese Speech Translation
Summary
This paper presents PiDA, a phonetically-informed data augmentation method for Vietnamese speech translation that improves robustness by generating ASR-like corruptions using phonetic word embeddings, achieving up to +2.04 BLEU on noisy outputs.
View Cached Full Text
Cached at: 06/12/26, 08:51 AM
# PiDA: Phonetically-Informed Data Augmentation for Robust Vietnamese Speech Translation Source: [https://arxiv.org/abs/2606.12911](https://arxiv.org/abs/2606.12911) [View PDF](https://arxiv.org/pdf/2606.12911) > Abstract:Cascaded speech translation \(ST\) systems suffer from error propagation when Automatic Speech Recognition \(ASR\) outputs incorrect transcripts\. We present the first systematic categorization of ASR errors for Vietnamese ST, classifying substitution errors by phonetic cause and quantifying their impact on downstream Neural Machine Translation \(NMT\) performance using Linear Mixed\-Effects Modelling\. We confirm that most ASR substitution errors arise from phonetic confusions rather than random noise, and that these phonetic errors significantly degrade ST quality\. Motivated by this finding, we propose Phonetically\-Informed Data Augmentation \(PiDA\), which generates ASR\-like corruptions by substituting words with phonetically similar alternatives using phonetic word embeddings\. Fine\-tuning on a PiDA\-augmented version of FLEURS Vietnamese\-English improves translation of erroneous ASR outputs \(up to \+2\.04 BLEU over standard fine\-tuning\) while also slightly improving clean\-text performance\. ## Submission history From: Giang Son Nguyen \[[view email](https://arxiv.org/show-email/15cfb280/2606.12911)\] **\[v1\]**Thu, 11 Jun 2026 05:09:59 UTC \(39 KB\)
Similar Articles
Phonetic Modeling of Dialectal Variation in Vietnamese Speech
This paper proposes a dialect-aware phonetic framework for modeling phonetic variation in Vietnamese ASR, decomposing syllables into structured components and mapping them to dialect-specific IPA representations. The approach matches pretrained baselines with fewer parameters and no external pretraining on the UIT-ViMD multi-dialect dataset.
BamiBERT: A New BERT-based Language Model for Vietnamese
BamiBERT is a new BERT-based pre-trained language model for Vietnamese that addresses limitations of PhoBERT, supporting longer context and operating without word segmentation, achieving state-of-the-art results on multiple Vietnamese benchmarks.
Low-resource Language Discrimination Towards Chinese Dialects with Transfer learning and Data Augmentation
The paper proposes a novel framework (CDDTLDA) using transfer learning and data augmentation to improve Chinese dialects discrimination under low-resource conditions, achieving state-of-the-art results on two benchmark corpora.
Contrastive Training with LLM-generated Near-Misses for Robust Code-Switching Speech Recognition
Proposes a POI-aware contrastive training framework using LLM-generated near-misses to improve ASR robustness at code-switching regions, achieving consistent error reductions on two benchmarks.
Towards a Phonology-Informed Evaluation of Multilingual TTS
This paper proposes a classifier-based framework to audit multilingual TTS systems for phonological faithfulness, using Assamese ATR vowel harmony as a case study. It reveals that Meta's MMS TTS frequently misproduces advanced tongue root vowels, a bias absent in human speech.