Error Rate is Not Enough: Costs of Augmentation in Low-Resource ASR
The paper argues that ASR augmentation can lower total error while worsening specific phoneme mistakes.
Across four low-resource languages and English, the authors compare audio-based and text-based augmentation methods. SpecAugment gives the steadiest overall gains, but raises tone and palatalization errors; text-based methods tend to add consonant insertions. Used together, the methods offset each other’s error patterns, with gains that are strongest in the low-resource setting and disappear by 100 hours of data. The authors recommend choosing augmentation based on training exposure and acoustic properties, measured before training. ArXiv · AI/CL/LG's note
Across four low-resource languages and English, the authors compare audio-based and text-based augmentation methods. SpecAugment gives the steadiest overall gains, but raises tone and palatalization errors; text-based methods tend to add consonant insertions. Used together, the methods offset each other’s error patterns, with gains that are strongest in the low-resource setting and disappear by 100 hours of data. The authors recommend choosing augmentation based on training exposure and acoustic properties, measured before training. ArXiv · AI/CL/LG's note
score 4