Transcription glossary

Automatic speech recognition (ASR)

Software that turns spoken audio into written text. Also called speech to text. Transcribe uses OpenAI's Whisper models for it, running on your PC.

Transcription

Writing down what was said in a recording. Automatic transcription is done by speech recognition software; the result is a draft that a person checks where it matters.

Offline transcription

Transcription that runs entirely on your own device, with no internet connection and no upload. The opposite of cloud transcription. Offline vs cloud.

On-device processing

Doing the computing on the device that holds the data, rather than sending the data to a server. It is what keeps recordings private in offline transcription.

Whisper

A family of speech recognition models released by OpenAI under the MIT licence, trained on a very large amount of multilingual audio. It writes punctuation and capital letters, detects the language, and handles accents and noise well. Whisper on Windows.

whisper.cpp

An open-source (MIT) program that runs Whisper models efficiently on ordinary computers, on the processor or on the graphics chip. Transcribe uses it as its engine.

Whisper small

A mid-sized Whisper model that balances speed and accuracy. Transcribe's Standard model, free without limits.

Whisper large-v3-turbo

A faster version of Whisper's largest model, with a much smaller decoder that keeps most of large-v3's accuracy. Transcribe's Accurate model.

Quantization

Storing a model's numbers with fewer bits so it takes less memory and runs faster, at a small cost in accuracy. Transcribe's models are quantized, which keeps the whole app at about 1 GB.

Word error rate (WER)

The standard measure of transcription accuracy: the words a transcript gets wrong (replaced, missing or added) divided by the number of words actually spoken. Lower is better; 2% means about two mistakes in a hundred words. Our measurements.

Real-time factor

How long transcription takes compared with the audio's length. "17× real time" means an hour of audio takes about three and a half minutes.

Signal-to-noise ratio (SNR)

How much louder the speech is than the background noise, in decibels. At 20 dB the noise is in the background; at 5 dB it is loud enough to make listeners strain.

Speaker diarization

Working out who spoke when: splitting a recording into turns and grouping the turns by voice, without knowing who the people are. Transcribe's speaker detection does this on your PC. Speaker identification explained.

Speaker identification

Recognising which known person is speaking. In everyday use it often means diarization plus naming: in Transcribe you name the speakers once and every line follows.

Timestamp

The time in the recording where a line or word begins. Timestamps tie a transcript to the audio, so you can jump to any moment and check it.

SRT

SubRip Text, the most widely supported subtitle format: numbered cues, each with a start and end time and the text to show. Make subtitles.

VTT (WebVTT)

Web Video Text Tracks, the subtitle format of the web. Like SRT, with a header and a few extra styling options; used by HTML5 video players and many platforms.

Loopback recording

Recording the sound a computer plays, such as the other side of an online call, directly from the sound system rather than through a microphone. Transcribe uses it, with your microphone, to record calls.

Live transcription

Transcribing while people are still speaking, so text appears a few seconds behind the speech. Transcribe does it when you record a meeting with Pro.

Voice activity detection (VAD)

Finding the parts of a recording that contain speech, so silence and noise can be skipped. It speeds up transcription and reduces text written during silence.

Hallucination (in speech recognition)

Text a model writes that was not said, most often during silence, music or noise. It is rare but possible with every Whisper-based tool, which is why important passages should be checked against the audio.

Custom words

Names and terms you give the engine in advance so it expects them. Also called a custom vocabulary or prompt. In Transcribe it is a Pro feature; on our names test it removed every error.

Verbatim transcript

A transcript that keeps every word as spoken, including repetitions and filler words. A clean verbatim transcript leaves fillers out. Whisper tends towards clean verbatim.

Vulkan

A standard way for programs to use the graphics chip for computing. Transcribe uses it to run speech recognition on graphics chips from AMD, Intel and NVIDIA, when their driver supports it.

Keep reading