Code switching

Evio

Moving between languages inside a single sentence, which is ordinary in Indian conversation and difficult for most systems.

What is code switching?

Code switching is the movement between two languages inside a single utterance, often inside a single clause. A caller asking about a loan in Hindi will routinely name the product and the amount in English without noticing the change. It is ordinary in Indian conversation and awkward for any system that expects one language per call.

Also called: Code-switching, Language mixing

Why a language setting is the wrong unit for an Indian call

Most voice platforms ask the deployment to declare a language for the call, which assumes callers hold to it. In Mumbai or Hyderabad that assumption fails in the first sentence, and it fails hardest on exactly the words the call depends on, because product names carry most of the intent alongside English numerals. A recogniser locked to Hindi transcribes English fragments as confident nonsense, and a recogniser locked to English does the reverse. Deployments that hold up treat mixed speech as the default case rather than as an exception to handle later.

  • Mixed speech is the normal register for urban Indian callers rather than an edge case.
  • The switched words are usually the ones carrying the intent or the amount.
  • A monolingual recogniser produces confident nonsense rather than an error it can flag.
  • An agent replying in mixed language is understood more readily than a formally correct one.

How code switching works

  • Call audio is sampled to establish which language pairs callers actually use.
  • A recogniser capable of mixed language decoding is selected over two separate models.
  • Vocabulary for product names and numerals is supplied in both scripts.
  • Intent detection operates on the mixed transcript rather than on a translation of it.
  • Responses are generated in the register the caller used, including the borrowed words.
  • Transcripts retain both languages so a reviewer sees what was actually said.

How code switching is evaluated

Code switching is assessed on the mixed utterances specifically, since an overall figure is dominated by monolingual speech. Measure the following on sampled live calls:

  • word error rate on utterances containing a language switch
  • error rate on the switched words themselves
  • intent accuracy on mixed utterances against monolingual ones
  • proportion of calls containing at least one switch
  • numeral accuracy where digits are spoken in the second language

Limitations, risks and common failure modes

  • Accuracy reported for Hindi is measured on Hindi that no caller actually speaks.
  • Two single language recognisers in parallel disagree, and arbitration loses the switch.
  • Translating into one language before intent detection discards the disambiguating words.
  • Replies in formal register alienate callers who opened in mixed speech.

Related terms

Barge-inAllowing a caller to interrupt the agent mid-sentence, as people do with each other.Dialogue managementThe logic that decides what the agent does next, given the state of the conversation and the caller's intent.Multi-turn conversationAn exchange where each reply depends on what came before rather than on the latest utterance alone.Turn takingManaging who speaks when, which is the hardest part of making a spoken exchange feel natural.Cold transferPassing a caller on with no context, which is where most of the frustration in handover comes from.Containment rateThe share of contacts resolved without reaching a person, which is the number most contact centres are judged on.

All 315 terms