Transcription Accuracy: How to Improve Results
What influences transcription accuracy and what you can do: engine choice, correct language, key terms, and recording quality.
Transcription accuracy depends on three factors: audio quality, the engine chosen, and the information you provide (language and key terms). This article shows what to adjust in each area to get the best result.
What influences the result most
- Audio quality — clear voices close to the microphone transcribe better than distant or muffled recordings.
- Engine choice — Premium is the high-accuracy standard; Standard handles noise best; Ultra delivers the most accurate speaker separation.
- Language selected — choosing the right language prevents the AI from "guessing" wrong. The platform supports over 160 languages and variants, with Brazilian and European Portuguese as a priority.
Adjustments you control when submitting
Set the correct language
If you know the audio's language, select it instead of leaving automatic detection on. Use "auto" only when you really don't know the language or when you receive varied files.
Use key terms (Premium engine)
On the Premium engine you can provide key terms — proper names, acronyms, technical jargon — that the AI will recognize with priority. It works very well in Portuguese. It's the best weapon against misspelled names of people and companies.
Choose the engine based on audio type
Clean audio: Premium. Audio with noise, strong accents, or difficult recordings: Standard. Overlapping voices where speaker attribution is critical: Ultra (Pro plans or higher). See the comparison in Choosing the engine.
Enable speaker detection when there's more than one voice
With detection active, the text comes out organized by speaker, with timestamps — which also makes it easier to review and correct. Details in Speaker identification.
Recording tips
- Record in a quiet environment with the microphone close to the speaker.
- Avoid two people talking at once — overlap is the hardest scenario for any AI.
- Prefer formats without aggressive compression when possible (the platform accepts and converts virtually any format, but already degraded audio limits the result).
Medical field: there is a Clinical Vocabulary in alpha phase that automatically injects drug, exam, and technical jargon terms. Access is granted upon request — see Custom vocabulary.
After transcription: review in the editor
No AI gets 100% right. For final adjustments, use the integrated editor: fix words, rename speakers, and save — exports (PDF, DOCX, subtitles) then use the corrected text. See Edit transcription.
If the result came far below expectations, first check the language configured at submission. Portuguese audio transcribed with the wrong language is the most common cause of a result that makes no sense.
Frequently asked questions
Is automatic language detection reliable?
Yes for most cases, but manually setting the language is always safer when you know it. Automatic detection is most useful when you process files from varied sources.
Do key terms work in Portuguese?
Yes. The feature is native to the Premium engine and works well in Brazilian Portuguese.
Which engine should I use for audio with a lot of noise?
The Standard. It was trained for difficult recordings and is still the cheapest in cycles. See Noisy audio.
Does correcting the transcription in the editor cost cycles?
No. Editing is free and unlimited — only the transcription itself consumes cycles.
Related articles
Didn't resolve? Open a ticket and our team will help you.