How to build a Hokkien Translator

About Hokkien

Hokkien (also known as Minnan, Taiwanese Hokkien, Taigi) is used by around 50 million speakers worldwide, more than Polish. It is relied upon throughout much of Southeast Asia; such as Singapore, Taiwan, Malaysia, China, and the Philippines. It is the most popular Min language. Like most Chinese languages, Hokkien is not mutually intelligible with other Chinese languages.

Hokkien is arguably the most important Chinese language for global companies after Mandarin Chinese and Cantonese Chinese. This is due to Hokkien’s international reach and speakers’ reliance on the language compared to other Chinese languages like Wu (Shanghainese). Because of its international reach, Hokkien is often found written in the Latin script, as not all Hokkien speakers understand the Chinese script.

Recommended Hokkien Translation Pipeline

Using English-Hokkien Translation as an example, the following Hokkien translation pipeline produces the best results:

Hokkien Variants

Most variants of Hokkien are mutually intelligible, but differences in grammar and vocabulary can occur by region. Whilst Taiwanese Hokkien is the most stable and recognized variant of Hokkien, Hokkien variants are also found outside of Taiwan. This includes Malaysia, Singapore, the Philippines, Hong Kong, Indonesia, Cambodia, China, Myanmar, Thailand, Brunei, and Vietnam.

Hokkien Written Scripts

There are 3 main ways of writing Hokkien:

• Pure Traditional Chinese Script (漢字; Hàn-jī)
• Mixed Script (Hàn-lô)
• Pure Latin Script (Lô-má-jī)

A standardized Traditional Chinese Script character set for Hokkien was released by Taiwan in 2007. Outside of Taiwan, Hokkien written in the Traditional Chinese Script remains non-standardized and can vary by community.

There are two types of Hokkien Latin Scripts: Tâi-lô and Pe̍h-ōe-jī (POJ). POJ is older but has more historic resources. Tailo is newer and officially adopted in Taiwan. Tailo is found in standardized uses of Hokkien such as in education and government. Both scripts are mutually intelligible, and can be converted between each other.

From a technical point of view, using Traditional Chinese Script as an intermediary helps produce better Hokkien translations. However, because a significant amount of the international Hokkien diaspora is not familiar with Chinese scripts, Latin Script is usually preferred when reading Hokkien. As such, to reach the most end-users, we recommend ensuring that the Latin Script is available. Translators can also choose to offer multiple scripts, for example; Microsoft offers Arabic in both Arabic and Latin/Roman scripts, whilst Google offers Crimean Tatar in both Latin and Cyrillic scripts).

Hokkien Text-to-Speech (TTS)

As of this writing, there are several ways to convert Hokkien text to speech (i.e. Hokkien TTS). Each of these come with their own set of considerations:

Hokkien Text-to-Speech	Rank*	Considerations
ITUAN (Taigi)	1	Can handle all formats (Pure Chinese Script, Mixed Script, Latin Script) Note: POJ appears to get converted to Tailo before TTS is performed
Facebook MMS (Chinese, Min Nan)	2	⚠️ Will not read Chinese script. Any Chinese characters must be converted to a Latin script before TTS. The differences between Tailo and POJ are minimal, though Tailo subjectively performs better.
Microsoft (Southern Min)	3	⚠️ Microsoft’s TTS voice does not pronounce Hokkien in Latin Script correctly, it’s not trained on it. As such, all Latin script portions need to be converted to Chinese script first. Despite Microsoft labeling their TTS Voice as Simplified Chinese script, it works fine with Traditional Chinese script. Whilst differences in pronunciation between the two scripts do exist, they are too small to be meaningful.

*Ranked by closeness to MoeDict Hokkien Human Recordings using the TTS tool’s optimal format (whether chinese script, mixed script, or latin script). 1=Best

Training a Hokkien Translation Model

Training a language translation model boils down to two steps:

1. Gather Data
2. Train a Model

Gathering Data

Most Hokkien datasets generally include English or Mandarin (in Traditional Chinese Script), as well as Hokkien in different written scripts (Traditional Chinese Script, Mixed Script, POJ Latin Script, Tailo Latin Script).

Click here for our list of [Hokkien Translation Datasets].

⚠️ Note: Most Hokkien Translation data is restricted to non-commercial use. If you are building a commercial translator check out our high quality Hokkien dataset that’s licensed for commercial use.

Training a Model

In 2025, one of the easiest ways to train a language translation model is to fine tune a large language model (LLM). OpenAI makes finetuning their base models very easy (see Finetuning OpenAI Models ). Meanwhile, with a bit more technical skill, one you finetune other models like LLAMA or do additional finetuning on existing Hokkien translation models (⚠️ Note: Existing Hokkien Translation Models are generally not trained on commercializable data and can not be used for commercial purposes).

Conclusions

We’ve presented various considerations when building a Hokkien translator. This ranges from the Written Scripts in the translation pipeline, to Hokkien variants and Hokkien Text-to-Speech. You should also consider if the translation model will be for non-commercial or commercial. If the translator is for commercial purposes, then clean high-quality commercial-use data is a necessity.

We look forward to seeing your Hokkien Translator out in the world soon!