Code-switching

What it is

In many markets, callers do not speak one language. An Indian caller may give an address in Hindi and a policy number in English, mid-sentence. A Filipino caller mixes Tagalog and English by default. Code-switching is that behaviour, and it is the normal register of business calls across South and Southeast Asia, much of Africa and many immigrant communities in the United States, United Kingdom and Europe.

Speech recognition is typically configured per language. A model set to Hindi renders English words phonetically in Devanagari, or drops them; a model set to English mangles the Hindi. Some engines offer a multilingual or code-switching mode; its accuracy on mixed speech varies widely and is rarely on the datasheet. Text-to-speech has the mirror problem: a voice built for one language reads the other with a heavy accent or wrong stress.

Why it matters when buying

If your callers code-switch and your vendor's demo was in clean English, you have not seen the product. Accuracy on names, numbers and place names in mixed speech is often the deciding factor for containment in India and the Philippines, and it drives which vendors are viable in a market at all. Language detection mid-call adds latency and can flip the wrong way on short answers.

What to ask

Ask which languages are supported in a single call and whether that means detection per turn or a genuinely multilingual model. Bring golden recordings of real mixed-language calls to the bake-off. Ask the agent to read back a Hindi name and an English policy number in the same sentence. Ask how the agent's own voice handles the second language.