Filipino and Cantonese: fil and yue codes in MAI-Transcribe-2 vs Sume

If you transcribe Filipino or Cantonese, MAI-Transcribe-2-Streaming expects the three-letter codes fil and yue that appear in its 60-language list, and it treats any unrecognized value as unset, which means auto-detect. Sume STT 1.0 takes language_code as a 2 to 16 character string and falls back to auto-detect when you omit it, but Sume does not publish a per-language support table, so the answer for a given language is a short test, not a lookup.

What the two request fields accept

On the Microsoft realtime page, transcription.language is null for auto-detect and the page lists the supported codes, including fil and yue among the three-letter ones. The same page says unrecognized values are treated as unset. That is convenient and also a trap: a code like fil-PH or yue-HK that is not on the list does not error, it quietly becomes auto-detect.

In the Sume request schema, language_code is a string between 2 and 16 characters described as a BCP-47 hint, and omitting it means auto-detect. The length range allows fil, yue and regional tags; whether the provider honors a specific code is something the schema does not promise.

A cheap check before you send

A shape check cannot tell you whether a language is supported, but it catches typos and stray region suffixes before they silently turn into auto-detect or an ignored hint. This one accepts two or three lowercase letters with an optional region and rejects the rest.

import re

TAG = re.compile(r"^[a-z]{2,3}(-[A-Z]{2})?$")

def check(tag):
    if not TAG.match(tag):
        raise ValueError(f"unexpected language tag: {tag!r}")
    return tag

for t in ["fil", "yue", "ko", "pt-BR", "fil_PH"]:
    try:
        print(check(t), "ok")
    except ValueError as e:
        print(e)

How to settle it for your language

Run a 30 second clip of real speech through Sume with your code, then once with the hint omitted, and compare the text and the returned language_code. See the stored walkthrough on testing a language with a sample for the routine. If the two outputs agree, you can omit the hint and let detection work.

Limits: one clip proves little for dialects. Cantonese spoken with heavy English mixing, or Filipino code-switched with English, is exactly where a single hint can hurt. Test with your own audio, not a clean read.

Before you commit to a language

A short routine saves a surprise later. None of it needs more than one clip per language.

• Record or pick 30 seconds of the speech you actually expect, not a studio read.
• Transcribe once with your hint and once without, and compare the text and the returned language code.
• Write down which of the two you ship, and the clip that justified it.
• Re-run when the vendor's language list or your audio source changes; the Microsoft page is a public preview and can change.