Code-Switching in Business Meetings: Why AI Transcription Apps Get It Wrong

What Is Code-Switching?

Code-switching is the practice of alternating between two or more languages within a single conversation, or within a single sentence. It's not sloppiness or laziness. It's a natural, rule-governed feature of bilingual and multilingual speech communities worldwide.

For Pakistani professionals, Urdu-English code-switching is the default register for educated business conversation. For Indian teams, it might be Hindi-English. For Moroccan companies, Darija-French or Darija-Spanish. For Singaporean teams, Singlish blends English, Malay, Mandarin, and Tamil in ways that feel completely natural to the speaker.

Linguists have studied this extensively: code-switching follows grammatical rules. Speakers don't mix randomly, they switch at specific syntactic boundaries, and the switches often signal something meaningful (like shifting from casual to formal, or emphasizing a technical term). Most AI models don't know any of this.

Why Mainstream AI Transcription Fails at Code-Switching

The problem is structural. A model trained on English audio learns to predict the next English word given the previous English words. When the speaker says something in Urdu, the model has two bad choices: guess the closest-sounding English word (producing nonsense), or skip the audio entirely (producing gaps).

Some more sophisticated models detect a language change and switch their internal decoding model, but this transition takes time, and in rapid code-switching (which happens at clause boundaries, often under half a second), the model is always catching up.

The practical result: a meeting transcript that's accurate for the English portions, garbled or empty for the Urdu portions, and missing the coherent bilingual meaning entirely.

• Words get dropped silently when the model doesn't recognize the phonemes
• Urdu words get phoneticized as English: 'kya' becomes 'kia' or 'kya'; 'ho gaya' becomes 'ho gaya' or disappears
• Sentences lose their meaning because the key verb or subject was in the other language
• Speaker diarization breaks as the model's language-switching mechanism disrupts its acoustic speaker model
• Action items assigned in Urdu never appear in the transcript at all

The Real Business Cost of Getting This Wrong

A garbled transcript isn't just annoying, it has measurable business consequences:

• Action items assigned in Urdu go unrecorded and don't get done
• Decisions made in the team's native language disappear from the official record
• Non-bilingual stakeholders (or overseas investors reading the transcript later) miss critical context
• Legal and compliance records are incomplete, which creates risk
• Teams waste time in follow-up meetings reconstructing what was decided
• Trust in the transcription tool erodes, and people stop using it, which is the worst outcome

What a Correct Solution Looks Like

Getting code-switching right requires training a model specifically on code-switched audio, not just concatenating monolingual corpora. This is technically harder because labeled code-switched training data is much rarer than monolingual English data.

The model also needs to handle script selection intelligently. When a speaker says 'kaam' (Urdu for 'work'), the output could be the Nastaliq کام, the Roman 'kaam', or even the English equivalent 'work'. A well-designed system lets the user decide, rather than making that choice silently and inconsistently.

Speaker diarization in code-switched audio needs special handling too. The acoustic signature of a speaker's voice changes slightly between languages, a naive diarizer can interpret a language switch as a new speaker entering the conversation, producing a mess of phantom speakers.

Samjha was built specifically for code-switched South Asian speech. The acoustic model is trained on real bilingual meeting audio, and the output gives you Urdu in both Nastaliq and Roman script with a toggle, so every speaker's contribution shows up correctly, in every language they used.

Practical Tips for Better Code-Switched Transcriptions

• Use a tool specifically trained on your language pair, not a generic English-first model with 'multilingual support' in the marketing copy
• Audio quality is the single biggest accuracy lever, a good headset mic beats everything else
• If your team speaks very fast (as many code-switched speakers do), consider a brief pause at key decision points
• Always review the transcript against the audio for high-stakes meetings, AI is very good but not perfect
• Use the AI chat feature to ask what was decided in plain language, rather than reading a potentially garbled transcript top-to-bottom
• Export and share in the script your audience reads most naturally, Nastaliq for formal records, Roman for quick team updates

The Bottom Line

Code-switching is not a problem to be solved, it's a communication style to be supported. The right tool doesn't ask your team to change how they speak. It meets them where they are, in the language they're using at that moment.

If your team regularly holds multilingual meetings and your current transcription tool is producing half a conversation, it's time to try something built for the way you actually talk.