AI Transcription Glossary: 40 Key Terms

AI transcription has its own vocabulary, and the same words show up whether you are comparing services, reading a benchmark, or debugging a messy transcript. This glossary defines the 40 terms that matter most in AI transcription and speaker identification — each in one or two plain-English sentences you can quote directly. Terms are grouped by topic: accuracy metrics, speech-recognition internals, audio and signal, speakers and diarization, output and formatting, and models and processing.

Quick Navigation

• Accuracy & Evaluation Metrics
• Speech Recognition Internals
• Audio & Signal Terms
• Speakers & Diarization
• Output, Formats & Timing
• Models & Processing
• Frequently Asked Questions

Accuracy & Evaluation Metrics

These are the numbers people cite when they claim one transcription system is more accurate than another. Every accuracy metric compares a system's output against a human-verified reference, so a metric is only as meaningful as the audio it was measured on.

Word Error Rate (WER) — The standard transcription accuracy metric: the count of substituted, deleted, and inserted words divided by the total number of words in the reference transcript. A 5% WER means 5 of every 100 words differ from the human reference, so lower is better.

Character Error Rate (CER) — The same idea as WER measured at the character level instead of the word level. CER is used for languages without clear word boundaries (such as Chinese and Japanese) and for judging near-miss errors like a single wrong letter.

Diarization Error Rate (DER) — The standard metric for speaker attribution: the percentage of audio time assigned to the wrong speaker, plus missed speech and false alarms. DER is independent of WER — a transcript can have accurate words but a poor DER if speakers are mislabeled.

Ground Truth — The human-created, verified transcript that a system's output is compared against to compute error rates. Accuracy claims are only valid relative to a specific ground-truth dataset.

Substitution, Insertion, and Deletion — The three error types that make up Word Error Rate: a substitution is a wrong word, an insertion is an extra word the system invented, and a deletion is a word it missed.

Confidence Score — A per-word or per-segment probability, usually 0 to 1, indicating how certain the model is about a given prediction. Low-confidence words are useful signposts for where to focus manual review.

For a deeper look at what these numbers really mean, see what actually determines transcription accuracy and our investigation of accuracy claims.

Speech Recognition Internals

These terms describe how an AI system turns sound into text. Modern engines chain several models together, and knowing the stages helps you understand where errors come from.

Automatic Speech Recognition (ASR) — The technology that converts spoken audio into written text; "speech-to-text" and "ASR" mean the same thing. Every transcription service is built on an ASR system at its core.

Acoustic Model — The component that maps raw audio features to sound units (phonemes or sub-word tokens). It handles accents, background noise, and audio quality.

Language Model — The component that predicts likely word sequences, so the system chooses "recognize speech" over the acoustically similar "wreck a nice beach." It supplies the grammar and context that make transcripts readable.

Decoding (Beam Search) — The process of searching many possible transcriptions and selecting the most probable one, rather than committing to the first guess. A wider search can improve accuracy at the cost of speed.

Tokenization — Splitting audio or text into the small units (tokens) a model actually processes. Sub-word tokenization lets a model handle rare words and names it never saw in training.

Inference — Running a trained model on new audio to produce a transcript, as opposed to training the model in the first place. Inference is what happens every time you upload a file.

Audio & Signal Terms

Transcription accuracy depends more on the source audio than on any other single factor, and these terms describe the recording properties that matter. Improving them at the recording stage is the cheapest way to improve a transcript.

Sample Rate — How many times per second the audio waveform is measured, in kilohertz (kHz). 16 kHz is the practical minimum for speech, and most transcription engines gain nothing above 16–48 kHz.

Bit Depth — The number of bits used to store each audio sample, which sets the dynamic range. 16-bit is standard for speech recording.

Bitrate — The amount of data used per second of compressed audio, in kbps; higher bitrate preserves more detail. MP3 at 128–192 kbps is a good balance for transcription.

Signal-to-Noise Ratio (SNR) — The level of the speaker's voice relative to background noise, measured in decibels. A high SNR (clear voice, quiet background) is the single strongest predictor of a clean transcript.

Codec — The format used to compress and store audio, such as MP3, AAC, or FLAC. Lossy codecs discard some data to save space; lossless codecs like FLAC keep all of it.

Clipping — Distortion that occurs when a recording is too loud and the waveform is cut off at its peaks, permanently destroying information the model cannot recover. Recording with headroom prevents it.

Our audio quality guide covers how to get these right when recording.

Speakers & Diarization

These terms concern the "who said what" problem, which is separate from getting the words right. BrassTranscripts performs this labeling automatically on every file.

Speaker Diarization — Partitioning audio by speaker to determine who spoke when, producing segments labeled Speaker 1, Speaker 2, and so on. It runs alongside transcription, not as part of it.

Speaker Identification — Attaching a known identity to a voice, as opposed to diarization, which only distinguishes anonymous speakers from one another. Most transcription workflows need diarization; identification requires a prior voice sample of each named person.

Speaker Embedding — A numeric fingerprint of a voice that lets the system tell speakers apart by comparing how similar two segments sound. Diarization clusters these embeddings to group each speaker's segments.

Clustering — The step that groups the voice embeddings in a recording into distinct speakers, and decides how many speakers are present. Similar-sounding voices are the hardest case for clustering.

Overlapping Speech — Moments when two or more people talk at once, which is the most common cause of diarization errors. Clean, non-overlapping turns produce far more reliable speaker labels.

Speaker Label — The generic tag (Speaker 1, Speaker 2) a diarization system assigns; you rename these to real names afterward. See how to replace Speaker 1 and 2 with real names.

For a fuller treatment, see what is speaker diarization and the speaker diarization models comparison.

Output, Formats & Timing

These terms describe what you actually download and how the words are tied to the audio timeline. BrassTranscripts outputs TXT, SRT, VTT, and JSON from a single upload.

Timestamp — A time marker attaching a word or segment to a point in the audio, written as hours:minutes:seconds. Timestamps make transcripts searchable and enable caption alignment.

Word-Level vs Segment-Level Timestamps — Word-level timing marks every individual word, while segment-level timing marks whole phrases or sentences. Word-level timing is needed for precise caption sync and audio-to-text editing; segment-level is enough for reading and search.

SRT (SubRip) — A widely supported caption file format that numbers each caption block with a start and end time. SRT is the default choice for uploading captions to YouTube and most video platforms.

VTT (WebVTT) — A caption format designed for the web (HTML5 video) that supports styling and positioning SRT lacks. Use VTT for websites and web players.

JSON Transcript — A structured, machine-readable transcript containing text, speakers, and word-level timing as data. JSON is the format to use when feeding transcripts into software or AI tools. See choosing the right transcript format.

Caption vs Subtitle — Captions are written for viewers who cannot hear the audio and include non-speech cues like [music]; subtitles assume the viewer can hear and often carry a translation. The underlying SRT or VTT file can serve either role.

Punctuation Restoration — The step that adds capitalization, commas, and sentence breaks to raw recognized words, since the acoustic model outputs an unpunctuated stream. Good punctuation restoration is much of what makes a transcript readable.

Verbatim — A transcription style that preserves every spoken word, including filler and false starts, as opposed to a cleaned-up version. See the verbatim vs clean verbatim guide for the full spectrum of styles.

Models & Processing

These terms come up when comparing how services are built and run. They explain the trade-offs behind speed, cost, and accuracy.

Whisper — OpenAI's open-source speech-recognition model family, widely used by developers and researchers as a transcription building block. It is a third-party model, not a service.

Fine-Tuning — Further training a general model on domain-specific data (for example, medical or legal vocabulary) to improve accuracy on that domain. It requires labeled data and compute, so most users rely on well-trained general models instead.

Batch vs Real-Time (Streaming) — Batch transcription processes a complete recording after the fact and can spend more computation per minute for higher accuracy; real-time transcription produces words live during a call but under tight time limits. BrassTranscripts is a batch service, which is why it prioritizes accuracy on recorded files.

Latency — The delay between speech and transcribed output, which matters for live captioning but not for uploaded files. Batch services trade latency for accuracy.

Hallucination — When a model outputs plausible words that were never actually spoken, often during silence, music, or noise. It is a known failure mode of neural transcription models and a reason to review low-confidence sections.

Diarization Pipeline — The chained sequence of voice-activity detection, embedding, and clustering that produces speaker labels. Each stage's errors compound, which is why clean audio matters so much for speaker accuracy.

GPU — The specialized processor that runs modern transcription models efficiently; self-hosting a model requires access to one. A managed service like BrassTranscripts removes that requirement by running the GPUs for you.

Frequently Asked Questions

What is Word Error Rate (WER) in transcription?

Word Error Rate is the standard accuracy metric for transcription, calculated as the number of substituted, deleted, and inserted words divided by the total words in the reference transcript. A WER of 5% means five of every 100 words differ from a human ground-truth transcript, so lower is better.

What is the difference between WER and DER?

Word Error Rate (WER) measures transcription accuracy — whether the words are correct — while Diarization Error Rate (DER) measures speaker-attribution accuracy — whether the right speaker was assigned to each segment. A transcript can have low WER but high DER if the words are right but the speaker labels are wrong.

What is speaker diarization?

Speaker diarization is the process of partitioning audio by speaker to answer "who spoke when," labeling each segment as Speaker 1, Speaker 2, and so on. It is separate from transcription: transcription produces the words, diarization attaches a speaker to them.

What is the difference between a caption and a subtitle?

Captions are a text version of speech intended for viewers who cannot hear the audio and often include non-speech cues like [music] or [applause], while subtitles assume the viewer can hear and typically render only dialogue, sometimes translated. In transcription workflows the SRT and VTT files BrassTranscripts produces can serve either purpose.

What does "verbatim" mean in transcription?

Verbatim transcription captures every spoken word exactly, including filler words, false starts, and repetitions, whereas clean or intelligent-verbatim transcription removes those disfluencies for readability. The right style depends on whether you need a faithful record or a polished document.

What is Voice Activity Detection (VAD)?

Voice Activity Detection is the step that identifies which portions of an audio file contain speech versus silence or noise, so the transcription engine only processes the parts that contain words. Good VAD improves both accuracy and processing speed by skipping non-speech audio.

Put the Terms to Work

Now that the vocabulary is clear, the practical questions are easier to answer: how to choose a transcription service, how audio quality drives accuracy, and which transcript format fits your workflow.

Ready to see it in practice? Upload a file to BrassTranscripts and get a speaker-labeled transcript in TXT, SRT, VTT, and JSON — no subscription required.