A friend was complaining: he bought some credit of iFlytek Machine Translate service, which claimed to have 95% accuracy rate translating English voice into Chinese subtitles. It was cheap – about US$0.09 for every minute of speech processed.

It was OK for a while, until he started translating a speech with a Singaporean speaker who speaks with Singlish accent. He estimated that the accuracy rate was probably at best 60%.

Well, to put it bluntly. If any AI system can claim to recognise and translate accented English accurately, I would sell myself to work for them.

Even in China, after so much effort and investment into voice recognition, we still manage to only capture only (standard) Mandarin and a few standardised version of dialects (such as Sichuanese).

The systems will usually not recognise deviations from standardised accents. Why? Because there are so many varieties of accents on top of the varieties of dialects, and it is not worth investing efforts to recognise them, as the number of speakers of each variation is limited.

You will not get much use case and economic benefits by catering to that type of voice recognition. It is much easier to teach each person standard mandarin.

For the same logic, it is much more worthwhile to invest in Bahasa Indonesia voice recognition than Singlish.

Besides, why should anyone complain about a service that costs only US$0.09 per minute?

As Jack Ma famously said (many years ago): if you bought a Rolex watch for CNY25 (US$3.9) on his ecommerce site, and complained that it was counterfeit. Whose problem is that? 


Thanks for reading The Low Down (TLD), the blog by the team at Momentum Works. Got a different perspective or have a burning opinion to share? Let us know at hello@mworks.asia.