Transcription and subtitle glossary

Words used in the app and guides

WER

Word error rate counts substitutions, deletions and insertions relative to a reference transcript. It is an evaluation measure, not a promise about the accuracy of a new recording.

CER

Character error rate uses characters rather than words as the evaluation units. It can be more appropriate for text that is not naturally separated into words by spaces.

Diarization

Diarization separates a recording into speaker turns. It does not necessarily establish a person’s identity. MurmurScribe offers manual speaker tags rather than automatic diarization.

Transcription

Transcription represents spoken language as written text. A useful draft preserves the intended wording, with uncertainty left visible until it can be checked.

Translation

Translation represents meaning in a different language. Whisper’s translation task produces English; it is distinct from writing the original spoken language.

Timestamp

A timestamp connects text to a point or interval in a recording. Segment timestamps help a reviewer return to the source audio.

Segment

A segment is a timed span of transcript text. Its boundaries are model output and may need review before the text becomes a caption cue.

Word timestamp

A word timestamp associates a timing interval with an individual word. This is more granular than the segment-level timestamps exposed by this app.

SRT

SubRip text is a plain subtitle format containing numbered cues, start and end times, and text. It is widely used for importing basic subtitles.

WebVTT

Web Video Text Tracks is a web caption format with a header and timed cues. It can include metadata and positioning that a plain-text conversion may discard.

Cue

A cue is a timed subtitle entry. Its start, end and text need to work together so the viewer can read the words while following the scene.

Caption

A caption communicates dialogue and relevant sound information. Accessibility captions may need speaker labels and meaningful sound descriptions in addition to speech.

Subtitle

A subtitle presents timed text with video, often dialogue or translated speech. Delivery rules determine what additional sound or speaker information is needed.

SDH

Subtitles for the Deaf and Hard of Hearing include relevant non-speech information and speaker cues. A speech-recognition draft is not a finished SDH deliverable.

Closed captions

Closed captions can be enabled or disabled by the viewer in a compatible player. They are distinct from words permanently rendered into the picture.

Open captions

Open captions are displayed as part of the video image. A separate SRT or VTT export does not by itself create an open-captioned video.

CPS

Characters per second divides the displayed text count by the cue duration. It is a reading-load check, not a complete measure of comprehension.

WPM

Words per minute expresses a rate of speaking or reading. A rate measured while reading a script should not automatically be applied to spontaneous conversation.

CPL

Characters per line is a subtitle layout measure. Different scripts, fonts and delivery formats can occupy different widths for the same character count.

Line break

A line break divides subtitle text visually. Breaks should preserve readable phrases rather than splitting tightly connected words solely to meet a count.

Reading speed

Subtitle reading speed describes the amount of text displayed during a cue. Dense concepts and unfamiliar names can be harder to read at the same numerical rate.

Offset

An offset moves every subtitle cue by the same time interval. It repairs a fixed timing difference, but cannot repair edits that change the timeline.

Drift

Drift is a timing difference that grows over a recording. A constant subtitle shift cannot correct it because the required adjustment changes with time.

Sample rate

Sample rate describes how frequently an audio signal is sampled. Resampling changes that representation; it does not recover detail absent from the original recording.

Bit depth

Bit depth describes the precision of individual PCM samples. It is different from the compressed bitrate used to estimate an MP3 file’s storage size.

Bitrate

Bitrate describes data used per unit of time. A compressed audio file’s size depends on its average bitrate and duration, with some additional container overhead.

Codec

A codec encodes and decodes audio or video. The browser must support the actual codec, not merely recognize a familiar filename extension.

Container

A container packages media tracks and metadata. MP4, WebM and Ogg can hold particular codecs; the container name does not fully specify the audio.

PCM

Pulse-code modulation represents a sampled signal using numerical values. Ordinary uncompressed WAV exports often contain PCM audio.

Lossless audio

Lossless compression preserves the encoded source samples. It also preserves noise and recording mistakes; it is not a repair process.

Lossy audio

Lossy compression reduces data by discarding information. Repeated encoding can add degradation, so retaining the original source is useful.

Mono

Mono audio has a single channel. MurmurScribe mixes decoded audio to mono for its speech model; separate speaker tracks are not retained for diarization.

Stereo

Stereo audio has separate channels. A meeting recording may place different sources in them, but channel count alone does not identify speakers.

Clipping

Clipping occurs when a signal exceeds the available recording range. Distorted peaks can obscure speech, and transcription cannot restore missing waveform information.

Noise floor

The noise floor is the underlying background level in a recording. Room noise and equipment noise can make quiet speech difficult to distinguish.

Reverberation

Reverberation is sound persisting through reflections in a room. A close microphone and a suitable recording space can help keep words distinct.

Normalization

Text normalization defines how written forms are simplified before scoring. Audio normalization adjusts levels. The same word describes different operations in these contexts.

Verbatim

A verbatim transcript aims to preserve the spoken wording. The treatment of fillers, repetitions and non-speech events should be defined for the project.

Clean verbatim

Clean verbatim removes selected disfluencies while preserving meaning. It needs a clear convention so editing does not silently change the speaker’s intent.

Filler word

A filler is a spoken hesitation or discourse marker. Removing it may suit a readable transcript but can be inappropriate for some forms of analysis.

Cross-talk

Cross-talk is overlapping speech. A model may omit a voice or combine the speakers, so those passages deserve direct listening.

Hallucination

A speech-model hallucination is text not supported by the audio. It can sound plausible and therefore survive casual proofreading unless checked against the recording.

Language detection

Language detection chooses a likely language from the recording. Short clips and mixed-language speech can make that decision uncertain.

Quantization

Quantization reduces the numerical precision used for model weights. It can reduce storage and memory demands, but its effect on results needs testing.

Model weights

Model weights are learned numerical parameters downloaded for inference. They are software data, separate from the recording selected by the user.

Inference

Inference is running a trained model on new input. In MurmurScribe, the browser performs this work locally rather than uploading audio to a remote inference service.

WebAssembly

WebAssembly is a browser execution technology used here for CPU inference. It provides a fallback when GPU processing is unavailable.

WebGPU

WebGPU exposes GPU capabilities to supported browsers. Availability depends on the browser, device and operating environment, and initialization can still fail.

IndexedDB

IndexedDB is browser-local structured storage. It can hold recordings and transcripts, but eviction, private browsing and site-data deletion can remove them.

Browser cache

A browser cache can retain downloaded model files for later use. Cached weights are not a promise that every part of the site will load while offline.

UTF-8

UTF-8 represents text using a standard character encoding. A plain subtitle file should retain the intended characters when moved between compatible applications.

JSON export

A JSON export preserves structured transcript fields for another workflow. MurmurScribe exports segment text, timestamps and manually entered speaker labels.

DOCX export

A DOCX export creates an editable document. It is useful for review and annotation, but it does not carry an interactive media player into the document.

FLEURS

FLEURS is a multilingual speech evaluation dataset used in the Whisper paper. Its published results describe that experiment, not every recording in a language.

Common Voice

Common Voice is a speech dataset used for another evaluation in the Whisper paper. Its results should be read separately from FLEURS because the recordings differ.

Sources

Whisper paper, appendix tables and normalization method · OpenAI / Radford et al. · Checked 2026-10-30
Digital audio concepts · MDN · Checked 2026-10-30
Browser audio codec guide · MDN · Checked 2026-10-30
WebVTT syntax and cues · MDN · Checked 2026-10-30
W3C prerecorded captions guidance · W3C WAI · Checked 2026-10-30
Browser model execution · Hugging Face · Checked 2026-10-30