Most quality assurance in modern contact centers runs on transcripts. Automated QA rules look for required phrases and banned words. Analysts search transcripts for product names, complaint themes, and competitor mentions. Dashboards count call reasons. All of this assumes that the same word is written the same way every time. In multilingual Southeast Asian contact centers, that assumption often fails — and the failure is silent.

Where do mixed scripts and spellings come from?

Script choice: in Thai, English loanwords can be written in Thai script (แพ็กเกจ) or Latin letters (package). In Singapore and Malaysia, Mandarin may be transcribed in Chinese characters or pinyin, alongside English. Spelling variation: Taglish affixed words appear as "i-cancel", "icancel", or "i cancel"; Indonesian informal words vary ("nggak", "gak", "enggak"). Diacritics: Vietnamese transcripts with and without tone marks. Product names: written in official form, phonetically, or abbreviated. Unicode: visually identical characters can have different underlying encodings, so a search for one doesn't match the other — the Unicode normalization forms exist to solve exactly this.

How does this break QA?

Quietly. A QA rule that checks for a required disclosure phrase misses calls where one word is spelled differently, and marks them as non-compliant — or a rule detecting a banned phrase misses variants and marks bad calls as fine. A keyword search for a product finds only some mentions. Trend dashboards undercount topics. Because nothing errors out, teams don't realize their QA and analytics are incomplete. Multilingual QA covers the wider QA programme.

What transcript conventions should you set?

For each language, write down: which script to use for loanwords and product names; standard spellings for common informal words and affixed forms; how numbers, amounts, and dates are written (digits or words); whether diacritics are always included; how to write brand and product names (official form); and Unicode normalization (usually NFC). Keep it short and practical, and apply it automatically in the transcription pipeline rather than relying on people.

How should conventions be enforced?

In the pipeline. Recognition output should be normalized automatically: Unicode normalization, script conversion for known loanwords, spelling standardization for common variants, and mapping of product name variants to official forms. Keep a list of variants per term, built from real transcripts, and update it as new products and slang appear. Where full normalization isn't possible, QA rules and searches should match on variant lists rather than single strings.

What about analytics across languages?

Map terms to shared concepts. "Cancel", "hủy", "ยกเลิก", "batal", and "i-cancel" are the same intent; product names in different scripts are the same product. Tagging transcripts with normalized intents and entities — rather than relying on raw words — makes cross-language reporting reliable. Voice-of-customer analytics covers how those insights are used.

How do you check whether transcripts are consistent?

Sample transcripts regularly and count variant spellings of your most important terms. If a product name appears in more than one form, or a required phrase has several spellings, add normalization rules. Test QA rules on known calls containing variants to make sure they're caught. Track this per language, since each has its own variation patterns.

How does WIZ.AI keep transcripts consistent?

Language Engine applies configurable transcript conventions per language — script choice for loanwords, spelling standardization, number formats, Unicode normalization, and product name mapping. QM runs rules and analytics on normalized transcripts and intents, so results are complete across languages and scripts.

Multilingual QA is only as good as the consistency of the text it reads. Book a demo to see QA on your languages.