Code-Switching Manifest

[SuperReem]

Hi,

Is this manifest configuration for the text field correct for code-switching fine-tuning? Also does language model training with aggregate tokenizer support this?

{
  "audio_filepath": "seg_36_360-45_870.wav",
  "duration": "9.509999999999998",
  "text": [
    {"lang": "ar", "str": "و"},
    {"lang": "ar", "str": "ايه"},
    {"lang": "en", "str": "I"},
    {"lang": "en", "str": "think"},
    {"lang": "ar", "str": "انه"},
    {"lang": "ar", "str": "رح"},
    {"lang": "ar", "str": "يكون"},
    {"lang": "ar", "str": "مشهد"},
    ]
}

[SuperReem]

@titu1994 @nithinraok

[KunalDhawan]

Hi @SuperReem,
Yes, the manifest is in accordance with the multi-sample LID format for code-switching (ref).
You might also want to check out online code switching-dataset functionality, if it fits your use case - #6579
Sorry, we currently don't support LM training with code-switched text + aggregate tokenizer.