Datasets:
Possible annotation inconsistencies in a few samples
#2
by Neural-Hacker - opened
First of all, thanks for releasing this dataset—it's been very useful for benchmarking Indian ASR models on Agriculture.
While evaluating a few STT models, I noticed that some reference transcripts don't seem to exactly match the corresponding audio. After listening to the audio multiple times, there appear to be a few cases where the annotation may contain minor transcription inaccuracies.
Since this dataset is intended as a benchmark, even small annotation differences can affect metrics like WER and lead to unfair comparisons between models.
Has anyone else observed similar cases? If it's helpful, I can share the sample IDs along with my observations so they can be reviewed.
Thanks!