Almost nobody in Manila speaks pure Tagalog for an hour. A sentence starts in Filipino, takes an English noun, puts a Tagalog affix on an English verb (nag-submit, i-check, magme-meeting) and finishes in English. Speech models are built to settle on one language per stretch of audio, and that is exactly what Taglish defeats.
The language is identified from the beginning of the file. If the recording opens with a long English introduction, the Tagalog that follows may be written as English-sounding nonsense or even translated; if it opens in Tagalog, English terms are usually kept but sometimes respelled. When a result looks wrong from the first line, trimming the file so it starts with clear Filipino speech and uploading again often helps. Either way, mixed passages are where your editing time will go.
There is much less transcribed Tagalog audio in the world than English or Spanish, and the models reflect that. A newscast, a scripted vlog or one clear speaker in a quiet room gives a draft that is mostly right. A barkada talking over each other, a classroom recorded from the back row or a call on a weak signal gives a draft with gaps and guesses. We do not promise a fixed accuracy for Filipino.
Common slips are easy to recognise once you know them: ng and nang swapped, the linker -ng attached or detached wrongly, reduplicated verbs shortened (pupunta written as punta), and particles such as po, nga, naman and daw dropped when spoken quickly. Hyphens in forms like mag-aral and pag-ibig are not always where a style guide wants them.
Filipino here means Tagalog-based national speech. Cebuano (Bisaya), Ilocano, Hiligaynon, Waray, Kapampangan and Bikol are separate languages, not accents, and a recording in any of them will be transcribed badly or forced into Tagalog words. Tagalog spoken with a Visayan or Ilocano accent is generally fine; Batangas and other southern Tagalog speech is recognised, with local words more likely to be missed.
Names take checking. Filipino surnames come from Spanish, Chinese and native sources, nicknames like Jun-Jun, Bong or Ging are everywhere, and place names such as Parañaque or Dasmariñas may lose the ñ. Slang, text-speak, jejemon spelling and gay lingo are largely outside what the model has learned. Numbers arrive in digits whether they were said in English, Spanish or Tagalog, which is worth verifying for prices and times.
Filipino uploads are usually thesis and research interviews, focus group discussions, online classes, barangay and office meetings, sermons, podcasts and YouTube vlogs. OPM songs and other music are not a good fit: lyrics sung over instruments transcribe poorly. As a guest you have 10 transcriptions of up to 3 hours each, with 5 hours in total. A free preview opens without an account, and a free account keeps the full transcript. AI chat with the recording, summaries and PDF or Word export are in the paid plans.
The audio is processed on our own servers with our own speech models and is not passed to a third-party speech API. For research interviews covered by an ethics protocol, that keeps the list of parties handling the data short.