Transcribe Welsh audio to text

Welsh model results

These are original Whisper paper FLEURS results, not tests of MurmurScribe or its quantized browser models. Lower error rates are better. The paper counts substitutions, deletions and insertions against the reference words after its normalization procedure.

Source: Whisper paper, Table 13 and Table 11. The datasets contain different recordings and scoring conditions. Read each column on its own; they are not interchangeable measurements of your audio.

Choose a starting model

Among the models offered here, small has the lowest published FLEURS WER for Welsh, at 59.3%. Small is 42.0 percentage points below Tiny on this test, calculated as 101.3 minus 59.3. This comparison is a starting point, not a prediction of your finished transcript.

Try Tiny to check browser compatibility, then compare the same spoken passage with the model suggested by the table. Keep the setting that gives you fewer corrections on your actual recording.

Training exposure and review

Figure 11 reports 73 hours of Welsh speech-recognition training audio. This is the printed value, not a new estimate; it does not count the separate translation panel. Training volume alone does not describe accents, domains or recording quality.

Set the audio language to Welsh when you know it. Keep the task on transcription to preserve the spoken language; translation produces English instead. Review names, borrowed words and changes of speaker against the recording. Export the corrected text only after checking passages you plan to quote. Keep uncertain phrases marked until a fluent listener can resolve them.