Many people think audio data collection for machine learning is a boring, mechanical process. In reality, it requires deep immersion in culture, language, and real user behavior. This is especially true in complex, multilingual regions like Southeast Asia and Malaysia.
Imagine this situation. A woman is standing at a checkout counter in a noisy supermarket, holding a crying baby. Suddenly her card gets blocked, and she calls the bank in panic.
Will a standard voice bot understand her through stress, background noise, and a crying child?
Or take another example. A conservative literature professor speaks to a bot in a highly formal tone, using the full legal names of banks. Meanwhile, students use slang, abbreviations, and switch between languages almost constantly.
This is where many global ASR models fail. 1. Code-switching and local slang In Malaysia, people naturally speak Manglish, a mix of Malay, English, and Chinese dialects. Standard models often lose context when they hear particles like “lah”, “leh”, “mah”, or “liao” at the end of sentences.
2. The gap between official language and real speech Models trained on clean online text expect phrases like “Malayan Banking Berhad” or “Kumpulan Wang Simpanan Pekerja”.
In real life, people say Maybank, CIMB, EPF, or KWSP.
They also use local expressions like “bank in” instead of “deposit money”. This is where many global systems start to break down.
3. Synthetic noise is not real life To save time, teams often add artificial noise to clean recordings. But models quickly adapt to those patterns.
Synthetic data cannot reproduce the true acoustic environment of Malaysia. Traffic, thunderstorms, food courts, crowds, and even cicadas can mask parts of speech in ways that matter for recognition quality.
What we see in Malaysia is not unique. The same patterns exist across Central Asia, Africa, and Latin America.
Anywhere people mix languages, use local slang, and interact with technology in noisy, real-life environments, global ASR models struggle.
How we approach this If you want accurate AI that works for real people, you have to collect data in real conditions.
We record in noisy environments. We ask speakers to express real emotions under stress. We work with local experts who understand how people actually speak.
We also integrate low-latency ASR directly into the collection pipeline so recordings can be checked and validated instantly.
Great customer experience starts with understanding. And for AI to truly understand people, training has to move beyond clean studios and into the real world.