Welsh transcription: an honest picture
Welsh is a low-resource language for speech models, with far less training data than English. The output is a draft that needs a full review. Expect corrections throughout, and check every name, place and technical term.
Clear audio makes the biggest difference. One speaker near the microphone in a quiet room, such as a lecture, a dictated text or a recorded interview, gives the most usable draft. A noisy meeting, a phone call, people talking over each other or speech over music gives much weaker output. There, the timestamps mostly help you find your place while you correct the text.
Correcting a draft is still faster than typing everything by hand. Play the file, follow the highlighted line, and click any line to hear it again. For minutes, subtitles or anything published under a Welsh language policy, ask a fluent Welsh speaker to read the final text. Mutations and missing accents are easy to overlook if you read quickly, and they are the errors readers notice first.
Welsh, English and auto-detect
Most Welsh speakers also speak English and switch between the two, often within a sentence. On auto-detect, a recording that opens in English, or a short clip, is likely to be treated as English, and the Welsh parts then come out as misheard English words. This page preselects Welsh to avoid that.
For bilingual recordings, choose the language that dominates the file. A council or school meeting held mostly in Welsh works best with Welsh selected; one held mostly in English, with Welsh in a few contributions, works better with English. In both cases the minority language will need more correcting.
Welsh spelling: points to check
These are the details a review of a Welsh transcript most often has to fix.
- Mutations: soft, nasal and aspirate mutations change the first letter of a word (Caerdydd, i Gaerdydd, yng Nghaerdydd). The model may write the unmutated form or the wrong mutation.
- Digraphs such as ll, dd, ch, ff, rh and th may be split, swapped or replaced with English-style spelling.
- The circumflex (to bach) on â, ê, î, ô, û, ŵ and ŷ is often missing, as in tŷ or dŵr.
- North and South Welsh vocabulary differs (rŵan and nawr, llefrith and llaeth). The model may write a word from the other region or a standard form.
- Place names and personal names need checking, especially longer ones.
- English loanwords common in everyday Welsh, such as parcio, ffonio or tecstio, may keep English spelling in one place and Welsh spelling in another, so make them consistent.
Who uploads Welsh recordings
Common uploads are lectures and seminars taught in Welsh, school and council meetings, chapel and church services, eisteddfod talks, journalist and research interviews, podcasts, and YouTube videos that need Welsh subtitles. Oral history projects record older first-language speakers, whose regional speech needs the most careful review.
Welsh learners use transcripts to follow along with recordings, and families outside Wales, including Welsh speakers in Patagonia, upload recordings of relatives. Patagonian recordings can mix Welsh with Spanish, so expect extra correction in those passages.
Uploading and exporting
WhatsApp voice notes, phone recordings and MP4 or MOV videos upload as they are. The transcript comes in timestamped paragraphs, and the player highlights the current line. Click a line to hear it again and fix the text in place.
Export TXT, DOCX, PDF, SRT, VTT or CSV, with timestamps and speaker names on or off. To read the content in English, use the Google Translate link on the transcript, or the AI prompt export with ChatGPT, Claude or Gemini. Speaker labels are available on paid plans or with a Minute Pack. Your files are never used to train AI models.