New task: speech recognition

[neubig]

Speech recognition is a standard generation task where the input is speech, output is text. For now, analysis could be done on the output side only.

• Evaluation metric: word error rate, character error rate
• Types of analysis: could be done on the level of individual deletions, substitutions, insertions. I have an example here: https://github.com/neubig/util-scripts/blob/master/counterrors.pl https://github.com/neubig/util-scripts/blob/master/error-diff.pl

[pfliu-nlp]

Sounds interesting: @neubig

• has some datasets of this task been covered by huggingface dataset?
• do we have en example of the system output file?

[neubig]

• Yes, here's an example: https://huggingface.co/datasets/librispeech_asr
• I'll work on getting system outputs