Indonesian exists in two registers that differ more than a learner expects. The formal standard, bahasa baku, is what a news reader, a lecturer at the podium or a minister uses. Everyday speech, shaped largely by Jakarta, replaces tidak with nggak or gak, sudah with udah, saja with aja and bagaimana with gimana, says gue and lo for I and you, and ends sentences with particles such as sih, dong, deh, kok and kan that carry tone more than meaning.
A transcript should record what was said, and that is the aim here: colloquial words stay colloquial and are not rewritten into the formal ones. Formal speech comes out cleanest, because it matches the written language the model has seen most. Casual speech is transcribed readably, with one caveat. Many colloquial words have no single agreed spelling, so nggak, gak and enggak may all appear in the same transcript. If the text is for a thesis appendix or a published interview, decide on one form and apply it with search and replace.
Indonesian builds words by attaching affixes to a root. From kerja, work, come bekerja, mengerjakan, dikerjakan, pekerja and pekerjaan. Colloquial speech has its own set: mengerjakan becomes ngerjain, and the suffix -in stands in for -kan and -i, as in bikinin or ambilin. Both systems are written as they were spoken. In fast speech a prefix can be swallowed, and membeli heard as beli changes little in meaning but does change the wording, which matters when you are quoting someone.
One spelling rule cannot be heard at all. Di as a passive prefix is attached to the verb, as in dibaca, while di as a preposition stands alone, as in di rumah. They sound the same, and Indonesian writers mix them up constantly; ke follows the same pattern. The model gets most of these right from grammar and makes the same slips people do. Repeated words such as anak-anak and orang-orang are normally written out with a hyphen, not with the figure 2 used in text messages.
Most Indonesians speak a regional language as well, and it enters their Indonesian. A speaker from Central Java drops in nggih, monggo or sampun; a speaker from Bandung adds the particles mah, teh and atuh. Short insertions like these are sometimes written correctly and sometimes turned into the nearest Indonesian word. A recording held mostly in Javanese or Sundanese is a recording in a different language, and it should not be expected to come out as usable Indonesian text.
English is the other ingredient, most famously in the South Jakarta habit of scattering which is, literally and basically through Indonesian sentences. Common English words and phrases are usually written in English. Where the language changes several times within a sentence, expect to correct more. Malay as spoken in Malaysia is closely related but is a separate standard with its own vocabulary: kereta is a car there and a train in Indonesia, percuma means free of charge there and useless here. A Malaysian recording will not read like Indonesian, and on such audio we do not promise the same quality.
Indonesian public life is dense with abbreviations. Some are read letter by letter, such as KTP and BPJS, and many are blends pronounced as words, such as puskesmas, Kemenkeu or ormas. Well-known ones are written as usual. Internal ones, the name of a campus unit, a company division or a local programme, are spelled by ear, and they are the first thing to fix. Religious talks add Arabic expressions such as insyaallah and alhamdulillah, whose Indonesian spelling varies from one writer to the next and may vary in the transcript too.
Personal names carry two spelling systems. Older names keep the pre-1972 spelling, so Soeharto and Suharto, Djoko and Joko sound identical, and the model cannot know which one a person uses. Amounts of money deserve a check as well. Rupiah figures run to juta, miliar and triliun, speakers shorten them in conversation, and the way digits and separators are written may not match Indonesian convention. Confirm prices and totals against the audio at the timestamp.
Typical Indonesian files are a recorded lecture or webinar, an interview for a skripsi or a tesis that has to be transcribed for the appendix, an office rapat saved from Zoom or Google Meet, a podcast episode, a sermon, or a long voice note forwarded from WhatsApp. Guests can transcribe 10 recordings of up to 3 hours each, and the free allowance is 5 hours in total across them. A free account does not add hours; it keeps the full transcripts and adds an editor and a history. Translation, AI summaries, AI chat and export to PDF or Word are in the paid plans.
Recordings are processed on our own servers with our own speech models and are not sent to a third-party speech API. For research interviews where the respondent was promised confidentiality, that is one fewer outside party to list.