In Vietnamese the pitch of a syllable decides which word it is, and the writing system records it. Ma is a ghost, má is mother or cheek, mà is but, mả is a grave, mã is a code and mạ is a rice seedling. On top of the tone marks come the letters that distinguish vowels, ă, â, ê, ô, ơ and ư, and the consonant đ, so a single syllable such as được or những can carry two marks at once.
Vietnamese typed without marks, the way people write in a hurry on a foreign keyboard, is a guessing game even for native readers. The transcript here is written with every mark from the first line, so it can go into a report, a thesis or a subtitle file without being retyped. One practical note for searching: the same word can be typed with the tone mark on a different vowel, hòa or hoà, depending on the convention the writer follows. If a search finds nothing, try the other placement.
Every Vietnamese syllable is written as a separate unit, but a great many words are two syllables long: sinh viên is a student, ngân hàng is a bank, máy tính is a computer. The space therefore tells you nothing about where a word ends, and neither does a pause in speech. The transcript follows normal Vietnamese writing, one syllable at a time, which is what a Vietnamese reader expects. Tools that count words or split text at spaces will count syllables instead.
For recognition this has a consequence. An unclear syllable cannot be repaired from the shape of a longer word around it, only from the meaning of the sentence. Clear speech gives clean text. In fast or overlapping speech the usual error is a real Vietnamese syllable with the wrong tone or the wrong final consonant, which looks correct at a glance and reads oddly in context. Those are easiest to catch by listening at the timestamp, not by proofreading the text alone.
Vietnamese has one spelling and three broad ways of pronouncing it. In Hanoi, tr and ch sound the same, so tranh, a picture, and chanh, a lime, are distinguished by context, and d, gi and r are all one sound. In the south the hỏi and ngã tones merge, so nghỉ, to rest, and nghĩ, to think, are pronounced alike, v is often said like a y sound, and several final consonants fall together. Vocabulary differs too: bố mẹ in the north is ba má in the south, and a thousand is nghìn in one and ngàn in the other.
The model writes standard spelling from whichever accent it hears and chooses between same-sounding spellings from context. Northern and southern speech of the kind heard in lectures, news and office meetings is recognised best. Central accents, Huế, Nghệ An or Quảng Nam for example, are difficult even for Vietnamese listeners from other regions, and a recording in strong central speech needs noticeably more correction. Regional words are kept as spoken and are not replaced with their northern equivalents.
A handful of family names, Nguyễn, Trần and Lê first among them, cover most of the population, so people are identified by their given name: anh Tuấn, chị Lan, cô Hạnh. Given names are where tone matters most, because Dung and Dũng are two different names, and a phone microphone in a meeting room does not always make the difference clear. Check names against a participant list. The kinship words that Vietnamese uses in place of you and I, such as anh, chị, em, cô and chú, are written as spoken.
Foreign words appear in two ways. Long-established borrowings are spelled as Vietnamese syllables, like cà phê and ô tô. Recent English terms from business and technology, marketing, deadline, online, are normally left in English, though a term said with a strong Vietnamese accent can come out as Vietnamese-looking syllables. Amounts of money need a check as well: speakers shorten them, as in một triệu rưỡi for one and a half million, and leave the unit to context.
Vietnamese uploads tend to be an hour or two long: a university lecture or an online class, a company meeting recorded in Zoom or Google Meet, an interview for research or for an article, a talk or a podcast on YouTube, or a string of long voice messages. Singing over music transcribes poorly in any language, so use this for speech. Guests get 10 transcriptions of up to 3 hours each, within 5 hours in total. A free account shares that same allowance; it keeps the full transcripts and adds an editor and a history. Translation, AI summaries, AI chat with the transcript and PDF or Word export belong to the paid plans.
The speech models run on our own servers. Your audio is not sent to a third-party speech API, which is worth knowing when the file is an internal meeting or an interview given in confidence.