Skip to content

BPE models - chunking of audio potentially "orphans" subwords of a word across chunks  #84

@leecming82

Description

@leecming82

I'm currently chunking up audio so that each chunk is a manageable duration using a BPE ASR model but I noticed there's the potential for a multi-subword word to be split across audio chunks and consequently, the subwords can get treated as separate words.

The pyctcdecode does the merging of subwords automagically so I'm wondering if there's a way to handle this edge case?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions