Transkripce
Česká verze

Automatic Transcription and Language Editing: Where the Machine Ends and the Human Begins

· 5 min read ·

Transcription is done. The model has returned text. Now what? The output of automation is not a finished document — it is a working version from which a different phase of work begins. Exactly where responsibility passes from machine to human, and why this boundary is not in the same place for every type of recording.

Two Phases of the Transcription Process

The traditional view of transcription as a single operation — listen and write — distorts the reality of the automated process. There are two distinct phases with different outputs and responsibilities.

Phase one: automatic transcription

The model captures the phonetic content of the recording. It receives the audio signal, decodes it through spectrograms and a neural network, selects the most probable word sequence, and returns text. It does this quickly, without fatigue, without distraction.

What it does not do: it does not interpret the speaker's intent. It does not correct grammar — it transcribes what it hears, not what would be correct. It does not decide about tone, irony, or cultural context. It does not know whether a statement is important or marginal.

Output of phase one: raw text, faithful to the audio, not always faithful to the meaning.

Phase two: human editing

The editor takes the raw text and works with it as material. They verify factual accuracy: proper names, numbers, quotations, specialist terminology. They perform language editing: grammar, punctuation, filler words as needed. They make interpretive decisions: what to preserve verbatim, how to formulate, what format of result matches the purpose.

Output of phase two: a document that serves its purpose.

What the Machine Captures Reliably

On a clean recording of standard conversational speech, the results of today's best models are good.

WER (word error rate) below 5% on professionally recorded standard speech is achievable for the best models. On one hour of recording (approximately 9,000 words) this means fewer than 450 problematic words — some of them minor spelling variants, gendered forms, or punctuation, not factually incorrect transcriptions.

For simple transcription — notes, a structured interview, a lecture with clear diction — the result is of good quality and editing takes less time than transcribing the recording from scratch.

Numbers and dates are transcribed phonetically correctly. Format depends on the language model: "one hundred and fifteen thousand" or "115,000". Editing is a matter of formatting, not content correction.

What the Machine Captures Poorly or Not at All

Homophones and Contextual Ambiguity

Language is full of phonetically similar or identical pairs: words that sound alike but differ in meaning. Context decides which is correct — the model handles this to varying degrees depending on how strong the contextual signal is.

In texts with a high frequency of homophonic words, the editor must go through every suspect point. Orientation: confidence score marks words where the model was uncertain — those are the first candidates for review.

Specialist Terminology Outside Training Data

The model does not know the term → it transcribes the phonetically nearest word from its vocabulary. The result looks grammatically correct and may pass unnoticed. This is the insidious nature of this error type: an unknown word is easy to spot, an existing-but-wrong word requires specialist knowledge to catch.

This is where the editor's competence plays the greatest role. An expert in the relevant field spots the error instantly. A generalist may miss it.

Intent, Tone, and Irony

"That is certainly an excellent solution." The model transcribes literally — with high confidence. Whether the sentence is praise or bitter irony depends on the context of the conversation, the tone of voice, and the relationship between the speakers. The model does not interpret tone. An editor who knows the context decides on the contextualisation of the quotation.

Logical Structure and Paragraphs

The model segments text according to acoustic pauses, not according to the logical construction of the statement. Result: segments too short or too long, paragraphs not corresponding to content. The editor reorganises structure to match content — this is an interpretive decision, not a technical fix.

How the Editor's Role Changes

Transcription automation does not change whether the editor works — it changes what they work on.

From transcribing to reviewing

A traditional transcriber listens to the recording and writes every word. One hundred percent of the work is generating text from scratch. Typing speed and auditory concentration are the key skills.

An editor in an automated process reads the finished text and compares it with the recording at points of doubt. Estimated time savings: 60 to 80% of original transcription time, depending on recording quality and text complexity.

New core competency: recognising model errors

The most valuable skill of an editor in an automation environment: knowing where the model failed. This requires two things — knowing the subject-matter context of the text and reading the transcript critically, not as a finished document.

A fast typist is not automatically a good transcript editor. A good editor is one who recognises the error — not one who types quickly.

When Editing Is Not Enough

For recordings with WER above 30%, editing is inefficient: it takes longer than manual transcription from scratch. Practical rule: if editing one paragraph takes longer than transcribing it from the recording, transcribe manually.

Causes of such poor transcription: poor recording quality, a strong accent or dialect, extremely specialised terminology without a terminology list. Solution: improve the recording (permanent effect, a single investment) or accept manual transcription as the better option for that specific recording.

Transcription automation does not challenge the existence of the editor. It only challenges where the editor's value lies. A transcriber who typed quickly becomes an editor who recognises model errors — and interprets, verifies, and decides on the final form of the document. For simple recordings this means a few minutes of work. For complex transcripts it means a shift of skills, not the disappearance of the work.

What an editor typically looks for in a journalistic interview transcript is described in the overview of transcription for journalism A20. How recording quality affects the volume of editing is explained in the audio preparation guide A12.

Sources: