Transcription and translation

Transcription is a solved problem you can have for nothing

Whisper, released openly by OpenAI, transcribes dozens of languages at a standard that was commercial-grade when it appeared, and it runs on ordinary hardware. whisper.cpp makes it fast on a CPU; faster-whisper is quicker again on a modest GPU. There is a hosted version, and paid services from several companies, and for most purposes the free local route is the correct answer.

The arithmetic that decides it: transcribing an hour of audio locally costs electricity and about five to fifteen minutes of your machine’s time. The same hour through a paid service costs a small amount of money and about a minute. If you transcribe occasionally, either is fine. If you transcribe daily, local is free and private.

What actually affects accuracy

Not the service, mostly. In rough order of impact:

Audio quality. A clip-on microphone beats every software choice. Recording quietly, close to the speaker, in a room without hard echo does more than any model upgrade.

Overlapping speech. Two people talking at once defeats everything. This is the main reason meeting transcripts contain nonsense.

Domain vocabulary. Medical terms, product names, local place names, people’s names. Most tools accept a hint list or prompt with expected terms, and supplying twenty words improves a transcript noticeably.

Accent and language. Performance is uneven and broadly tracks how much of that speech was in training. Test on your own recordings rather than trusting a published word error rate.

Speaker separation. Knowing who said what is a separate problem from what was said, called diarisation. It is markedly less reliable than transcription itself, and it is where paid services add value: pyannote is the main free option and it takes work to set up.

Translation: three different tools

Dedicated translation systems. DeepL and Google Translate. Fast, cheap or free, strong on common language pairs, and generally better than a general assistant at plain translation of ordinary prose.

General assistants. Better where context and judgement matter: a legal clause, a joke, a message that must land in a particular register, or anything where you can explain the situation. Worse at bulk.

Open translation models. NLLB-200 from Meta covers 200 languages; MADLAD-400 from Google covers over 400. Both run locally and both matter enormously for languages the commercial services handle badly or not at all. Argos Translate packages offline translation for the desktop.

The workflow that beats each of them alone

For anything important: translate with a dedicated system first, then ask a general assistant to review the translation against the original, with the context explained. The second pass catches register errors, idioms taken literally, and terms that are technically correct and wrong in the setting.

For long documents, translate section by section. Quality degrades over long inputs, and a section boundary is a natural place to check.

Getting a transcript you can use

Two settings improve most transcripts more than any model choice. Ask for timestamps, so you can jump to the audio at the point where a number or a name appears and verify it in ten seconds rather than listening again from the start. And ask for the output as plain text rather than subtitle files if you intend to read it: subtitle formats break sentences at line lengths, which makes a transcript hard to read and harder to summarise.

whisper meeting.m4a --model medium --language en --output_format txt

The model size is the one real trade-off: larger models are more accurate and slower, and on a laptop the medium size is usually the sensible middle.

The limits worth stating plainly

Machine translation into a language you cannot read is a risk you are taking, not a task you have completed. For anything consequential (a contract, medical instructions, a safety notice) a human speaker needs to read it. This is not caution for its own sake: the failure mode is fluent, confident and undetectable from the side you understand.

Transcription errors are silent. A transcript reads perfectly while containing a wrong number, a wrong drug name or a negation dropped. Where the words matter, listen to the audio at the points that matter rather than trusting the text.

Recording other people has rules, and they differ by country and by setting. That is the subject of a later lesson in this module, and it applies before any of this does.