AI Transcription for Singapore Accents and Mixed-Language Calls

Why Singlish and code-switching trip up transcription

Speech recognition models learn from huge amounts of recorded speech. Most of it is American and British English, so a Singapore accent is less familiar, and local words such as lah, can or not and tabao often come out as nonsense or get dropped.

The bigger problem is code-switching: changing language inside one sentence ("Can you send the 报价单 by Friday?"). Many tools ask you to choose one language per meeting, then force everything into it. Mandarin phrases become English-sounding gibberish, or the reverse.

Singapore has public work on this. IMDA's National Speech Corpus is a large set of Singapore English recordings for researchers, and A*STAR's MERaLiON model is described by IMDA as handling English, Mandarin and Singlish "with advanced code-switching abilities" (IMDA annex, January 2026). These are research and developer tools, not ready-made meeting apps, but they show the problem is being worked on locally.

Which AI transcription tools handle mixed languages

This table summarises what each vendor's own documentation says, as of October 2026. We have not run our own accuracy tests, so treat it as a list of capabilities to check, not a ranking.

Sources: Teams multilingual speech recognition, Google Meet transcript languages, Otter language support, Otter pricing and OpenAI API pricing. Otter's help page says it "can only transcribe in one language at a time", so set the language before each meeting.

For a wider comparison of meeting note takers, including summaries and bots, see our guide to AI meeting notes tools and the PDPA.

Built-in meeting transcription vs a transcription API

There are two ways to get transcripts, and they suit different businesses.

If you are still choosing a meeting platform, our Zoom vs Teams vs Google Meet comparison covers the wider trade-offs.

Settings that improve accuracy for Singapore speakers

You cannot fix an accent, but you can give the software a better chance.

1. 1

   Set the right language before the meeting

   Tools that work one language at a time (Otter, Google Meet) guess badly if left on the wrong setting. In Teams with multilingual speech recognition, ask each speaker to set their spoken language.

2. 2

   List the expected languages for API transcription

   OpenAI's speech-to-text guide says gpt-transcribe takes a `languages` list (for example English and Chinese) instead of a single language.

3. 3

   Add keywords and context

   Feed in names, product codes, street names and jargon as keywords, plus a one-line description of the call. OpenAI notes keywords are hints, so check they do not appear where nobody said them.

4. 4

   Use a decent microphone

   A laptop mic at the end of a meeting room table is the most common cause of bad transcripts. A cheap conference speakerphone in the middle of the table helps more than switching software.

5. 5

   One speaker at a time

   Overlapping talk confuses every tool. If you need to know who said what, use a speaker-labelling (diarization) option rather than guessing later.

6. 6

   Split long recordings sensibly

   OpenAI's file API accepts files up to 25 MB. For long recordings, compress the audio or split at pauses, not mid-sentence.

PDPA: consent, notice and where recordings are stored

A recording or transcript of identifiable people is personal data under Singapore's PDPA. The practical duties are the same as for any other personal data; see PDPC's summary of the data protection obligations. This is general information, not legal advice.

• Notify and get consent: say at the start that the meeting is being recorded and transcribed, and why. Put the same line in the calendar invite for external guests.
• Purpose: use transcripts for what you said (meeting notes, quality checks), not for unrelated things such as staff performance scoring without telling them.
• Protection: keep transcripts in your company account, not personal Otter or ChatGPT accounts, with MFA on.
• Retention: delete recordings and transcripts when they are no longer needed. Set an auto-delete period if the tool has one.
• Overseas transfer: most AI vendors process audio outside Singapore. Check the vendor's data processing terms before you send customer calls.

On the API side, OpenAI's enterprise privacy page says it does not train models on business and API data by default. Consumer accounts can have different settings, which is one more reason to keep work audio off personal accounts.

A 30-minute accuracy check before you rely on a tool

Vendor claims are a starting point. A short test on your own audio tells you more than any review, because it uses your speakers, your room and your vocabulary.

Do it yourself or get help?

If your meetings are mostly in English, turn on the transcription already built into your meeting platform, add a recording notice to your invites and run the accuracy check above. That costs nothing extra.

If you transcribe mixed-language calls or recorded files every week, an automated workflow saves time: recordings land in a folder, get transcribed with your languages and keywords, and the transcript and action items go to the right person. Our Workflow Automation service (S$390 per workflow) sets this up in your own accounts, with retention and access set from day one.