When a conversation switches languages, keep listening.
What code-switching means for recognition, translation, and the interface around them.
Humlet editorialPeople often reach for the phrase that feels most natural, even when it comes from another language. They might explain a problem in one language, use a familiar workplace term in another, and switch back before finishing the sentence.
This is code-switching. For a speech product, it means the language setting at the start of a session may not describe everything that follows.
Treat language switching as ordinary speech.
A mixed-language sentence is not necessarily a translation request. The speaker may want the original words transcribed exactly as spoken. Another reader may want the same message translated into a shared language. Those are two distinct outputs.
Preserve that distinction in the interface. A transcription view should not quietly translate unfamiliar words away, and a translation view should leave the source available for checking.
Look inside the turn.
Testing one English recording followed by one Malay recording is useful, but it does not cover a sentence that combines them. Include language changes within a single utterance and across short exchanges.
Names and borrowed terms deserve special attention. A product name does not always signal a change in the whole sentence’s language. Evaluate whether the result preserves the term while keeping the surrounding message coherent.
Use context carefully.
Expected languages and specialist vocabulary can provide useful context when supported by the service. They should not become an excuse to discard unexpected speech. A conversation can include a visitor, a quoted phrase, or a language the application did not anticipate.
Record which hints you supplied during evaluation. A system given the right answer in advance is being tested under a different condition from one asked to detect the language automatically.
Design around continuity.
Humlet’s approach to speech routing starts with continuity: an application should be able to follow a conversation as its language changes. The proposed integration layer focuses on recognition and translation as connected but distinct tasks.
For now, the directory and prepared playground help explain the choices. A live integration needs to verify the language transitions, source text, translated results, and recovery behavior on real recordings. The most persuasive proof is a conversation a listener recognizes as normal.
A change of language should not mean starting the conversation again.
Building something that listens?
Explore the tools, or prepare a brief for your next speech integration.