What is WER, and what score is good enough?

Word Error Rate is the count of substituted, omitted and inserted words divided by the number of words in the reference text. Below 10% a transcript is comfortable to read; above 20% it needs a full pass by hand and costs more time than it saves.

Does automatic transcription work well outside English?

Yes, since the wave of large multilingual models. Whisper, released by OpenAI in September 2022 and trained on 680,000 hours of audio, marked a step change for non-English languages. English still leads, but the gap has narrowed sharply.

What is speaker diarization?

It is working out who speaks when in a multi-voice recording. It is handled separately from transcription and decides whether an interview or a panel is readable at all: without it, two speakers contradicting each other merge into one block of text.

What is audio-to-text transcription?

Audio-to-text transcription is the automatic conversion of recorded speech into written text by a speech recognition system. Its quality is measured by word error rate (WER), which now falls below 10% on clean audio in most major languages.

How does it actually work?

Audio-to-text transcription, or automatic speech recognition, is the conversion of a spoken audio signal into written text by software. It is the mandatory step for treating a podcast, a video or a meeting the way you treat an article: searching inside it, pulling quotations from it, summarising it.

The field is old. The Audrey system, built at Bell Labs in 1952, recognised the digits zero through nine spoken by a single speaker. The decisive shift is far more recent: abandoning component-based acoustic models in favour of end-to-end trained neural networks, which became the norm during the 2010s.

How is transcript quality measured?

The standard indicator is WER, Word Error Rate, computed as the sum of substitutions, omissions and insertions divided by the number of reference words. A WER of 8% means eight wrong words per hundred expected.

WER	Readability	Viable use
Under 5%	Near perfect	Direct quotation, subtitles
5 to 10%	Good	Reading, automatic summarization, search
10 to 20%	Degraded	Keyword search only
Over 20%	Poor	Full manual review required

WER has a known flaw: it treats every error as equivalent. Confusing "their" and "there" costs the same as botching a proper noun or a figure, even though the second wrecks the value of the passage. Which is why serious evaluation looks separately at accuracy on named entities and on numbers.

What degrades the result?

Capture quality. A lavalier mic and a phone mic in a reverberant room do not produce the same WER, model held constant. This is the dominant factor, ahead of model choice.
Overlapping speech. Two people talking at once produce an error almost every time, whatever the system.
Specialist vocabulary. Technical terms, acronyms, rare proper nouns: the words carrying the most information are the worst recognised, for lack of training occurrences.
Accents and speaking rate. A strong regional accent or a very fast speaker raises the error rate noticeably.
Language switching. A speaker dropping an English sentence into another language trips up models not designed for code-switching.

What changed after 2022?

OpenAI’s release of Whisper in September 2022 was the tipping point for languages other than English. The model was trained on 680,000 hours of multilingual audio collected from the web, an unprecedented volume in open access, and shipped under a permissive licence in several sizes. The immediate effect was to make decent-quality transcription available without a cloud vendor contract.

Competition then moved onto three fronts: latency for real-time use, diarization — knowing who is speaking — and word-level timestamping, which is what lets a quotation point back to its exact moment in the recording.

What do you do with a raw transcript?

A raw transcript is barely readable: a wall of text, spoken in syntax, full of repetitions and false starts that go unnoticed while listening and jump off the page in writing. It only has value as the input to a further step.

Three treatments change everything: splitting into paragraphs with punctuation restored, attributing turns of speech, and timestamping. With those three, an hour of podcast becomes a document you can consult, summarise and quote to the second. Without them, it stays a file nobody opens.

The order matters too. Punctuation should be restored before summarisation, because a model given an unsegmented block infers sentence boundaries in places the speaker never paused — and every downstream claim inherits that mistake.

Where does transcription fit in a monitoring workflow?

Transcription is not an end in itself: it is what brings audio and video into the same processing regime as text. Once transcribed, a podcast can be summarised, indexed, quoted and connected to the articles you read on the same subject — none of which is possible while it remains an audio file.

That is the position transcription holds in Synthiz: YouTube videos and podcasts you add are transcribed, then synthesised and poured into a thematic "memory" on the same footing as an article or a PDF, with quotations tied to the source passage. An idea heard in one episode is then found by searching, rather than by half-remembering which episode it was in.