OOD-Speech: A Large Bengali Speech Recognition Dataset for Out-of-Distribution Benchmarking
Paper • 2305.09688 • Published • 1
Datasets:
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('csv', {}), NamedSplit('test'): ('webdataset', {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Bengali Speech Recognition
Transcribe short Bengali utterances. Training clips carry transcripts; held-out clips do not, and their references are not in this dataset. Audio is 16 kHz mono mp3 at 48 kbps.
Contents
The data tree lives under data/, exactly as the benchmark environment presents it at
/app/bnspeech. Its own description is at data/README.md.
- 5 files, 1916446809 bytes (1.8 GiB)
- structure hash (sha256 over sorted
path\tsizelines):4f0422e65a3547d42c09276d3e8cf1621eb20e7af1737bfd23993b12c072560d data.manifest.tsvat the repo root lists every file assha256 size path
Use
from huggingface_hub import snapshot_download
snapshot_download("Emulated-Inc/bnspeech", repo_type="dataset", revision="REVISION",
local_dir="./bnspeech", allow_patterns="data/*")
Pin revision to a commit sha rather than a branch if you need reproducibility.
Provenance and licence
Derived from the Bengali.AI Speech Recognition Kaggle competition training corpus (https://www.kaggle.com/competitions/bengaliai-speech), released under CC BY 4.0. Attribution to Bengali.AI and the OOD-Speech dataset authors (Rakib et al., "OOD-Speech: A Large Bengali Speech Recognition Dataset for Out-of-Distribution Benchmarking", Interspeech 2023, arXiv:2305.09688) and the Mozilla Common Voice Bengali contributors. This is a curated, re-encoded subset with opaque ids and normalized transcripts; the held-out split is carved from the competition's public training data.
- Downloads last month
- 54