Identify the behavior you need
Distinguish occasional borrowed terms, proper names, short quotations, and sustained switches between languages. These behaviors create different recognition and language-identification challenges. Describe the target telephone workflow and provide examples so a source can explain whether relevant mixed-language material is actually present.
Choose annotation granularity
Decide whether language tags belong to calls, turns, utterances, or individual words. Include an unknown or ambiguous category where the speech does not support a confident decision. Preserve the underlying audio so reviewers can inspect transitions rather than accepting a transcript’s spelling choices as definitive language evidence.
Keep transcription rules compatible
Agree on scripts, transliteration, product names, and words commonly used across languages. A consistent rule matters more than silently switching conventions between annotators. State whether the evaluation expects the spoken language as written, a normalized representation, or a separate translation, and keep those outputs clearly separated.
Request representative transition cases
Inspect switches near interruptions, overlap, short responses, and domain terminology. Report coverage by language combination and switching pattern rather than only total multilingual hours. Share examples and reference requirements with Fiund so possible sources can be reviewed for project fit, sensitive content, and permissible annotation scope before further planning. For a transition review, select an exchange where the caller changes languages while the agent responds in the original language. Ask reviewers to label the same interval independently. Their differences can reveal whether the schema supports mixed speech and uncertain boundaries or is implicitly forcing each turn to have one unambiguous language.
Review checklist
- Define switching patterns and combinations
- Choose tag granularity and uncertainty rules
- Specify script and transliteration policy
- Inspect actual transition segments