Glossary
364 terms used across the wiki, each linked to the articles that explain them.
A
- Abstractive summarisation
- Building a summary by generating new sentences that paraphrase and compress the source, usually with a language model.
→ Conversation summarization - Accent
- A pattern of pronunciation shared by speakers from a region, social group or first language.
→ Accents and regional speech - Accent gap
- The difference in error rate a speech recognition system shows between speaker groups with different accents on comparable audio.
→ European English - Acoustic echo cancellation (AEC)
- Signal processing that removes the agent's own playback from the microphone signal so it is not mistaken for user speech.
→ Turn-taking and barge-in→ Automotive voice recognition - Acoustic model
- The component that scores how well a stretch of audio matches a given sound unit or sequence of words.
→ Acoustic models and language models - Adequacy decision
- A European Commission decision that a third country, or a framework within it, ensures a level of data protection essentially equivalent to the EU's.
→ EU data residency for speech AI - Agent assist
- Real-time software that listens to a live call and surfaces prompts, knowledge articles or checklists to the human agent.
→ Contact centre speech analytics - AI scribe
- A common name for software that produces ambient clinical documentation.
→ Ambient clinical documentation - Air-gapped deployment
- An installation with no network connection to the internet or the vendor, updated only by transferring signed artefacts manually.
→ On-premise and VPC deployment - Alphanumeric
- A string that mixes letters and digits, such as a postcode, reference number or registration plate.
→ Numbers and alphanumerics - Ambient clinical documentation
- Recording a clinical consultation and using speech recognition and a language model to draft the clinical note for the clinician to review.
→ Ambient clinical documentation - Anti-context test set
- Evaluation audio that contains none of the bias terms, used to check that biasing does not damage ordinary recognition.
→ Over-biasing - ASR hallucination
- Fluent text in a transcript that has no basis in the audio, as opposed to a mis-hearing of words that were actually spoken.
→ Hallucinations in speech recognition - Attention encoder-decoder (AED)
- An architecture in which a decoder generates text token by token while attending over the full sequence of encoder outputs.
→ ASR model architectures - Attribution error
- An error in which a statement is assigned to the wrong speaker, for example recording a relative's symptom as the patient's.
→ Ambient clinical documentation - Audio intelligence
- The set of techniques that extract structured information from audio beyond a verbatim transcript, such as speakers, topics, sentiment, entities and non-speech events.
→ What is audio intelligence? - Audio redaction
- Replacing the stretch of a recording that contains personal data with silence or a tone, located using word timestamps.
→ PII redaction in transcripts - Audio tagging
- Assigning one or more sound-class labels to a whole clip, without saying when each sound occurs.
→ Audio event detection - Audio token
- A discrete unit produced by a neural audio codec, used to represent speech as a sequence a language model can predict.
→ Speech-to-speech models vs pipelines - Audio-language model
- A model that takes audio directly as input and produces text outputs such as answers, summaries or labels, without a separate transcription step.
→ What is audio intelligence? - Automatic speech recognition (ASR)
- The task of converting spoken audio into written text by machine.
→ What is speech recognition? - Autoregressive decoder
- A decoder that predicts each output token conditioned on the tokens it has already produced.
→ Hallucinations in speech recognition
B
- Back-end speech recognition
- A dictation workflow in which the recording is transcribed automatically and a medical transcriptionist edits the draft before the clinician signs it.
→ Medical speech recognition - Backchannel
- A short listener response such as "mm-hm" or "yeah" that signals attention without claiming the turn.
→ Turn-taking and barge-in - Backpressure
- Slowing or buffering the sender when the receiver or network cannot keep up, so that queues do not grow without bound.
→ Integrating a speech-to-text API - Barge-in
- A user speaking over a system while it is still talking, which the system must detect and usually respond to by stopping.
→ What is voice AI?→ VAD vs endpointing vs turn detection→ What is a voice agent?→ Turn-taking and barge-in - Batch (pre-recorded) transcription
- Transcription of a complete audio file after recording has finished, returning the full transcript when processing is done.
→ Real-time vs batch transcription - Batch size
- The number of audio segments processed together in one pass through the model.
→ Real-time factor (RTF and RTFx) - Beam search
- A decoding strategy that keeps the few most probable partial transcripts at each step instead of committing to a single best guess.
→ How speech-to-text works→ Custom vocabulary and context biasing - Beamforming
- Combining signals from several microphones to reinforce sound from one direction and suppress sound from others.
→ Transcribing noisy audio→ Automotive voice recognition - Benchmark contamination
- The presence of test-set material, whether audio, transcripts or their sources, in the data a model was trained or tuned on.
→ Benchmark contamination - Bias list
- A set of words or phrases the recogniser should favour, typically names, products or technical terms.
→ Long-context biasing - Biased-word WER
- Word error rate computed only over the words in a biasing or keyword list, reported alongside WER on all other words.
→ Entity and keyword accuracy - Billing increment
- The unit to which each request's audio duration is rounded up when it is charged, such as one second or fifteen seconds.
→ Speech-to-text pricing models - Biometric data
- Personal data resulting from specific technical processing of physical, physiological or behavioural characteristics that allows or confirms the unique identification of a person (GDPR Article 4(14)).
→ GDPR and voice data - Boost weight
- The score bonus given to a listed term during decoding; higher weights make the term more likely to win against similar-sounding alternatives.
→ Custom vocabulary and context biasing→ Over-biasing - Brazilian Portuguese (BP)
- The Portuguese of Brazil, spoken by the large majority of Portuguese speakers and dominant in public speech data.
→ Portuguese speech recognition
C
- Calibration
- The agreement between stated confidence and observed accuracy: of words scored 0.8, a well-calibrated system gets about 80% right.
→ Confidence scores in speech recognition - Cascaded pipeline
- A voice system built from separate models chained together, typically speech-to-text, a language model and text-to-speech.
→ What is voice AI?→ What is a voice agent?→ Voice agent architecture→ What is audio intelligence? - Catastrophic forgetting
- The loss of previously learned ability when a neural network is trained further on a narrow new dataset.
→ Fine-tuning vs context biasing - Certification scope
- The part of an organisation, its services and locations, that the ISMS and its certificate actually cover.
→ ISO 27001 for speech AI vendors - Character error rate (CER)
- The number of character substitutions, deletions and insertions needed to turn a transcript into its reference, divided by the number of characters in the reference.
→ Character error rate (CER) - Characters per second (CPS)
- A cue's text length divided by its duration, used to check that viewers have time to read it.
→ Captions and subtitles: SRT, VTT, and timing - Check digit
- A character computed from the rest of an identifier so that most single-character errors can be detected.
→ Transcribing structured identifiers - Chunk
- A short block of audio, typically 20 to 100 ms, sent to or processed by a streaming recogniser in one step.
→ Streaming speech recognition - Chunking
- Splitting long audio into shorter segments that a model can process, either at fixed intervals or at detected pauses.
→ Long-form transcription - Clean verbatim
- A transcript that keeps the full meaning but removes fillers, stutters and false starts for readability.
→ Legal transcription - Clitic placement
- The position of unstressed object pronouns relative to the verb, which differs between European (diz-me) and Brazilian (me diz) Portuguese.
→ Portuguese speech recognition - Closed compound
- A compound written as a single word with no space or hyphen, such as German Krankenhaus or Dutch zorgverzekering.
→ Compound words and morphology - Closed-set LID
- Language identification where the answer is assumed to be one of a fixed list of languages.
→ Spoken language identification - Cloud ASR
- Speech recognition run on a provider's servers, reached over the network through an API.
→ On-device vs cloud speech recognition - Code-switching
- Alternating between two or more languages within one conversation, sentence or phrase.
→ Code-switching in speech recognition - Codec
- An algorithm that encodes and decodes audio, usually compressing it; distinct from the container file format that holds the encoded stream.
→ Audio formats for speech recognition - Compound splitting
- Breaking a closed compound such as Rechtsschutzversicherung into its parts, used in scoring and language modelling.
→ German speech recognition - Compression ratio
- The ratio of a transcript's raw length to its compressed length, used as a cheap signal of repetitive, looping output.
→ Hallucinations in speech recognition - Concurrency
- The number of requests or streams a system is handling at the same moment.
→ Concurrency and scaling - Conditional independence assumption
- CTC's assumption that each frame's output is independent of the others given the audio, which makes it fast but blind to what it has already written.
→ ASR model architectures - Confidence score
- A number, usually between 0 and 1, that a recogniser attaches to a word or utterance as an estimate of how likely it is to be correct.
→ Confidence scores in speech recognition - Content data vs metadata
- Content is the audio and transcript themselves; metadata is information about a request, such as its ID, duration and timestamps.
→ Zero data retention - Context biasing
- Supplying domain words or text to a speech model at inference time so that decoding favours them, without changing the model's weights.
→ Fine-tuning vs context biasing→ Long-context biasing - Context window
- The maximum amount of text, measured in tokens, that a language model can read in one request.
→ Using transcripts with LLMs - Controller
- The organisation that determines the purposes and means of processing personal data (GDPR Article 4(7)).
→ Data processing agreements - Controller and processor
- The controller decides why and how personal data is processed; the processor processes it on the controller's behalf and instructions.
→ GDPR and voice data - Conversational AI
- Systems that hold a multi-turn dialogue with a person, through text, voice or both.
→ Voice AI vs speech AI vs conversational AI - cpWER
- Concatenated minimum-permutation WER: WER computed per speaker after matching hypothesis speakers to reference speakers, so it penalises both recognition and attribution errors.
→ Speech recognition evaluation beyond WER - Credit-based pricing
- A model where plans include a pool of credits and each unit of usage, such as a minute of streaming or batch audio, consumes a fixed number of them.
→ Speech-to-text pricing models - Critical error
- A transcription error that changes clinical meaning, such as a wrong drug, dose, side or negation.
→ Medical speech recognition - Cross-lingual transfer
- The improvement a language gets from training data in other languages, typically largest for languages with little data of their own.
→ Multilingual speech recognition - CTC (connectionist temporal classification)
- A training objective and decoding scheme in which a network emits a token or a blank symbol for every frame, and repeats and blanks are collapsed to produce text.
→ ASR model architectures - CUDA graph
- A recorded sequence of GPU operations that can be replayed with one launch, removing per-kernel CPU overhead.
→ GPU inference for speech models - Cue
- A single caption or subtitle block with a start time, an end time and one or two lines of text.
→ Captions and subtitles: SRT, VTT, and timing - Curse of multilinguality
- The observation that, at fixed model size, adding more languages eventually lowers accuracy per language because capacity is shared.
→ Multilingual speech recognition - Custom vocabulary
- A list of words or phrases supplied to a speech recognition system so that it is more likely to output them.
→ Custom vocabulary and context biasing
D
- Data processing agreement (DPA)
- The contract between a controller and a processor that GDPR Article 28(3) requires before the processor handles personal data on the controller's behalf.
→ Data processing agreements - Data protection impact assessment (DPIA)
- An assessment required by GDPR Article 35 before processing likely to result in a high risk to people's rights and freedoms.
→ GDPR and voice data - Data residency
- A commitment about the geographic location where data is stored and processed, usually expressed as a region or set of countries.
→ EU data residency for speech AI - Deployer
- The AI Act's term for an organisation that uses an AI system under its authority, as distinct from the provider that develops it or places it on the market.
→ The EU AI Act and speech recognition - Deposition
- Sworn out-of-court testimony recorded for later use in litigation, common in common-law jurisdictions such as the United States.
→ Legal transcription - Diacritic accent (tilde diacrítica)
- A written accent that distinguishes otherwise identical words, such as tú (you) and tu (your).
→ Spanish speech recognition - Dialect
- A regional or social variety of a language that differs in vocabulary and grammar as well as pronunciation.
→ Accents and regional speech - Dialogue manager
- The component that tracks the state of a conversation and decides the system's next action.
→ Voice AI vs speech AI vs conversational AI - Diarization error rate (DER)
- The share of reference speech time that is missed, falsely detected, or attributed to the wrong speaker.
→ What is speaker diarization?→ Speech recognition evaluation beyond WER - Diglossia
- A situation where speakers use one language variety for everyday speech and another for writing and formal settings, as with Swiss German and Standard German.
→ German speech recognition - Distractor
- A term in a biasing list that is not spoken in a given utterance, and which the system may wrongly insert.
→ Custom vocabulary and context biasing→ Long-context biasing - Domain adaptation
- Changing how a speech recognition system behaves so that it performs well on a specific kind of audio, vocabulary or phrasing that differs from its general training data.
→ What is domain adaptation in ASR? - Domain mismatch
- The difference between the data a model was trained on and the data it is used on, in acoustics, vocabulary or style.
→ What is domain adaptation in ASR? - Domain prior
- Knowledge of which words and phrases are likely in a given domain, supplied to the recogniser to shift its choices.
→ Text-only domain adaptation - DTMF
- Dual-tone multi-frequency signalling: the tones produced by telephone keypad presses, carried in-band or as RTP events.
→ Voice agents on the phone - Dual-channel recording
- A call recording that stores each party on its own audio channel.
→ Telephony transcription→ Contact centre speech analytics - Dynamic batching
- Grouping requests that arrive close together into one batch at serving time, trading a small wait for higher throughput.
→ GPU inference for speech models - Dynamic time warping (DTW)
- An algorithm that aligns two sequences spoken at different speeds by stretching and compressing time, used in early template-matching recognisers.
→ A brief history of speech recognition
E
- E-set
- The group of English letter names that differ only in their initial consonant (B, C, D, E, G, P, T, V, Z) and are easily confused by recognisers.
→ Transcribing structured identifiers - Early cut-off
- An endpoint fired while the speaker still intended to continue, truncating the utterance.
→ What is endpoint detection? - EBU-TT-D
- The European Broadcasting Union's XML subtitle format for distribution, based on W3C TTML.
→ Media and broadcast transcription - Edge deployment
- Running a model on hardware close to where data is produced, such as a device or an on-site server, rather than in a central data centre.
→ On-device vs cloud speech recognition - Effective cost per hour
- Total spend on transcription divided by hours of audio actually processed, after rounding, minimums, add-ons and unused commitment.
→ Speech-to-text pricing models - Elision
- Dropping a final vowel before a vowel-initial word and marking it with an apostrophe, as in l'hôpital or j'ai.
→ French speech recognition - Emotion recognition system
- Under the AI Act, an AI system that identifies or infers people's emotions or intentions on the basis of their biometric data.
→ The EU AI Act and speech recognition - EN 301 549
- The European standard setting accessibility requirements for ICT products and services, referenced by the Web Accessibility Directive.
→ Speech recognition in the public sector - Encoder
- The part of a speech model that turns a sequence of acoustic frames into a sequence of context-aware vectors.
→ How speech-to-text works - End-of-query detection
- A model that predicts, from acoustic and decoder features, that a spoken query has ended.
→ What is endpoint detection? - End-of-turn detection
- Deciding, in real time, that a speaker has finished their conversational turn and expects a response.
→ End-of-turn detection - End-to-end model
- A single neural network trained to map audio features directly to text, replacing separately trained acoustic, pronunciation and language models.
→ What is speech recognition? - Endpoint
- The point at which a recogniser decides an utterance is complete and finalises its transcript.
→ What is endpoint detection? - English as a lingua franca (ELF)
- English used as a shared working language between speakers who do not have English as their first language.
→ European English - Enrolment
- Recording reference speech from a known person and storing a voiceprint for later comparison.
→ Diarization vs identification vs verification - Entity recall
- The share of names or key terms in the reference that appear correctly in the transcript.
→ Rare words and proper nouns - Equal error rate (EER)
- The operating point at which a verification system's false acceptance rate equals its false rejection rate.
→ Speech recognition vs voice recognition→ Diarization vs identification vs verification→ Voice biometrics and consent - Error propagation
- The way mistakes made by an early stage, such as a misrecognised name, carry through to every stage that depends on it.
→ What is audio intelligence? - EU AI Act
- Regulation (EU) 2024/1689, the EU's horizontal law on artificial intelligence, which sets obligations according to the risk an AI system's use poses.
→ The EU AI Act and speech recognition - European Case Law Identifier (ECLI)
- A standard identifier for court decisions in the EU, made up of the prefix ECLI, a country code, a court code, a year and an ordinal number.
→ Legal transcription - European Portuguese (EP)
- The Portuguese of Portugal, also the written reference for most African Portuguese-speaking countries.
→ Portuguese speech recognition - Exact-match rate
- The share of identifiers, such as order numbers or plates, transcribed with every character correct.
→ Entity and keyword accuracy - Expected calibration error (ECE)
- The weighted average gap between mean confidence and accuracy across confidence bins.
→ Confidence scores in speech recognition - Explicit consent
- Consent given by a clear, express statement, the standard GDPR Article 9(2)(a) requires for processing special category data.
→ Voice biometrics and consent - Exponential backoff with jitter
- A retry policy that waits a randomised, exponentially growing interval between attempts to avoid synchronised retry storms.
→ Integrating a speech-to-text API - Extractive summarisation
- Building a summary by selecting sentences or spans that already appear in the source text.
→ Conversation summarization
F
- Faithfulness
- Whether every statement in a summary is supported by the source conversation.
→ Conversation summarization - False accept
- A detection fired when the keyword was not spoken, usually reported per hour of audio.
→ Keyword spotting and wake words - False barge-in
- An interruption triggered by something other than the user deliberately taking the turn, such as echo, background noise or a backchannel.
→ Turn-taking and barge-in - False insertion (biasing)
- A bias term that appears in a transcript even though the speaker did not say it.
→ Over-biasing - False reject
- A missed detection when the keyword was spoken, usually reported as a percentage of true occurrences.
→ Keyword spotting and wake words - Far-field speech
- Speech captured by a microphone at a distance from the talker, so that room reflections and noise make up a large share of the signal.
→ Transcribing noisy audio - Final devoicing
- The pronunciation of a voiced consonant as voiceless at the end of a word, so that Polish chleb (bread) ends in a /p/ sound.
→ Polish speech recognition - Final latency (TTFS)
- The delay between the end of speech and the moment the final transcript for that speech reaches the client.
→ Speech-to-text latency - Final result
- A transcript segment the recogniser has committed to and will not change.
→ Streaming speech recognition→ Partial vs final results - Final-transcript latency
- The time from the end of speech to the arrival of the final transcript for that speech, the figure that gates a voice agent's reply.
→ Streaming speech recognition - Fine-tuning
- Continuing to train a pretrained model on domain-specific data so that its weights change to fit that domain.
→ Fine-tuning vs context biasing - Flemish
- The Dutch spoken in Flanders, the northern part of Belgium; it shares the standard written language with the Netherlands but differs in pronunciation and some vocabulary.
→ Dutch speech recognition - FLEURS
- A multilingual speech benchmark of read sentences in 102 languages, built on the same n-way parallel text in every language.
→ ASR benchmarks and leaderboards→ Speech recognition for European languages - Forced alignment
- Finding the time position of each word in a known transcript by aligning it against the audio with an acoustic model.
→ Word timestamps and forced alignment - Format constraint
- A description of what a valid identifier looks like, typically a regular expression, used to guide or check recognition.
→ Transcribing structured identifiers - Frame
- A short window of audio, typically 25 ms long taken every 10 ms, from which one feature vector is computed.
→ How speech-to-text works→ What is voice activity detection (VAD)?→ Voice agent frameworks - Frame rate (encoder)
- The time step of the model's output frames, commonly 20 to 40 ms, which bounds the precision of timestamps.
→ Word timestamps and forced alignment - Front-end speech recognition
- A dictation workflow in which the clinician sees the transcript as they speak and corrects it themselves before signing.
→ Medical speech recognition - Full-duplex model
- A model that listens and speaks at the same time, modelling both sides of a conversation as parallel streams.
→ Speech-to-speech models vs pipelines
G
- G.711
- The ITU-T codec standard for 8 kHz, 64 kbit/s telephone audio, in μ-law and A-law variants.
→ Voice agents on the phone - Gaussian mixture model (GMM)
- A weighted sum of Gaussian distributions, used in HMM systems to model the acoustic features emitted by each state.
→ A brief history of speech recognition - Gemination
- Lengthening a consonant so that it contrasts with its short counterpart, as in Italian pala (shovel) and palla (ball).
→ Italian speech recognition - Grammatical case
- A form of a noun or adjective that marks its role in the sentence; Polish has seven.
→ Polish speech recognition - Grapheme-to-phoneme (G2P)
- The step that converts spelled words into the sequence of speech sounds used to pronounce them.
→ What is text-to-speech?
H
- Half-cascade
- A voice agent design where an audio-native language model consumes speech directly but emits text for a separate text-to-speech system.
→ Speech-to-speech models vs pipelines - Hangover
- A period after speech ends during which the detector keeps reporting speech, so that brief dips and trailing consonants are not cut off.
→ What is voice activity detection (VAD)? - High-resource language
- A language with large amounts of transcribed speech and text available for training, such as English.
→ Speech recognition for European languages - High-risk AI system
- An AI system that is a safety component of a regulated product or is used in one of the areas listed in Annex III of the AI Act, and which must meet requirements such as risk management and human oversight.
→ The EU AI Act and speech recognition - Hybrid ASR
- An arrangement in which an embedded recogniser handles commands in the vehicle and a cloud recogniser handles open-ended requests when connected.
→ Automotive voice recognition→ On-device vs cloud speech recognition - Hypothesis
- The transcript produced by the speech recognition system under evaluation.
→ What is word error rate (WER)?
I
- In-domain evaluation
- Testing a model on data drawn from the same distribution as its training data, such as a corpus's test split after training on its training split.
→ Benchmark contamination - Indirect prompt injection
- Instructions hidden in data a model reads, such as a tool result or retrieved document, that attempt to change the model's behaviour.
→ Tool calling in voice agents - Inflection
- Changing the form of a word to mark grammatical features such as case, number, gender or tense, as Polish does with nouns, adjectives, verbs and numerals.
→ Polish speech recognition - Information security management system (ISMS)
- The set of policies, processes, risk assessments and controls an organisation uses to manage information security.
→ ISO 27001 for speech AI vendors - Inter-annotator agreement
- How closely two independent transcribers agree on the same audio, often expressed as the WER of one against the other.
→ Building your own ASR test set - Internal language model
- The implicit model of word sequences that an end-to-end ASR system learns from the transcripts of its training audio.
→ Text-only domain adaptation - Intra-sentential code-switching
- Changing language within a single sentence, as in a Dutch sentence containing English technical terms.
→ Multilingual voice agents - Intra-sentential switching
- A language change inside a single sentence, as opposed to between sentences.
→ Code-switching in speech recognition - Inverse real-time factor (RTFx)
- Audio duration divided by processing time; how many seconds of audio are processed per second of compute.
→ Real-time factor (RTF and RTFx) - Inverse text normalisation (ITN)
- Converting the spoken form of words produced by a recogniser into their conventional written form, such as 'twenty euros' into €20.
→ Punctuation, casing, and inverse text normalization - ISIN
- International Securities Identification Number: a 12-character code (ISO 6166) made of a two-letter country code, nine alphanumeric characters and a check digit.
→ Speech recognition in financial services - ISO/IEC 27001
- The international standard that specifies requirements for establishing, operating and continually improving an information security management system.
→ ISO 27001 for speech AI vendors
K
- Kernel
- A function that runs on the GPU; each layer of a model is executed as one or more kernels.
→ GPU inference for speech models - Kernel fusion
- Combining several GPU operations into one kernel so intermediate results stay in fast on-chip memory.
→ GPU inference for speech models - Keyword precision
- The share of target-term occurrences in the transcript that were actually spoken, which falls when a system inserts terms that were not said.
→ Entity and keyword accuracy - Keyword recall
- The share of target-term occurrences in the reference that appear correctly in the transcript.
→ Entity and keyword accuracy - Keyword spotting (KWS)
- Detecting when one of a small set of predefined words or phrases occurs in an audio stream.
→ Keyword spotting and wake words - Knowledge distillation
- Training a smaller student model to reproduce the outputs of a larger teacher model.
→ Model size and quantization
L
- L1 transfer
- The influence of a speaker's first language on the sounds, stress and grammar they use in a second language.
→ European English - L2 speech
- Speech in a language that is not the speaker's first language, typically carrying features of their first language.
→ Accents and regional speech - Language depth
- How well a speech system handles a language beyond basic word recognition: formatting, regional varieties, domain vocabulary and evaluation in that language.
→ Speech recognition for European languages - Language embedding
- A fixed-length vector computed from speech, trained so that utterances in the same language lie close together.
→ Spoken language identification - Language model
- A model of which word sequences are likely in a language or domain, independent of any audio.
→ Acoustic models and language models - Language routing
- Selecting the recognition language, prompt and text-to-speech voice for a conversation based on a detected or declared language.
→ Multilingual voice agents - Language tag
- A standard code identifying a language and optionally a region, such as nl-BE or fr-CH, defined by BCP 47.
→ Multilingual voice agents - Language token
- A special input or output symbol that tells a multilingual model which language to transcribe, or records which one it detected.
→ Multilingual speech recognition - Latency budget
- An allocation of the total acceptable response delay across each stage of a system, used to decide where optimisation effort goes.
→ Voice agent latency budget - Liaison
- In French, pronouncing a normally silent final consonant when the next word begins with a vowel, as in les amis said with a /z/.
→ French speech recognition - Linear PCM
- Uncompressed digital audio in which each sample is stored as an integer or float proportional to the signal amplitude.
→ Audio formats for speech recognition - Linking morpheme
- A sound or letter inserted between the parts of a compound, such as the -s- in arbeidsovereenkomst or the -en- in pannenkoek.
→ Dutch speech recognition - Little's law
- In a stable system, the average number of items in it equals the arrival rate multiplied by the average time each item spends there (L = λW).
→ Concurrency and scaling - LLM-as-judge
- Using a language model with a rubric to grade transcripts or outputs, as a scalable substitute for some human review.
→ Testing voice agents - Loanword
- A word borrowed from another language and used as part of the host language, often with adapted pronunciation or inflection.
→ Code-switching in speech recognition - Locale conventions
- The rules a language region uses for writing numbers, dates, currency and punctuation.
→ Speech recognition for European languages - Log-mel spectrogram
- A time-frequency picture of audio on a perceptual frequency scale, the standard input to modern speech models.
→ What is speech recognition? - Lombard effect
- The involuntary tendency of speakers to raise their voice and change its pitch and spectrum when talking in noise.
→ Automotive voice recognition - Long tail
- The large set of individually rare words, such as names and technical terms, that together make up a meaningful share of speech in a specialist domain.
→ What is domain adaptation in ASR?→ Rare words and proper nouns - Long-context biasing
- Context biasing that accepts large amounts of free text, such as whole documents, instead of a short list of phrases.
→ Long-context biasing - Long-form transcription
- Transcribing recordings much longer than a model's native input window, typically minutes to hours, by segmenting and recombining the audio.
→ Long-form transcription - Look-alike, sound-alike (LASA) drugs
- Medication names that are easily confused because they are spelled or pronounced similarly, such as hydroxyzine and hydralazine.
→ Medical speech recognition - Lookahead
- How much future audio a streaming model may see before committing to text for the current moment.
→ Real-time vs batch transcription→ Speech-to-text latency
M
- Managed voice platform
- A hosted service that runs voice agents for you, configured through a dashboard or API rather than your own orchestration code.
→ Voice agent frameworks - Map-reduce summarisation
- Summarising a long input in chunks, then summarising the chunk summaries into one result.
→ Conversation summarization - Match error rate (MER)
- The share of aligned word pairs that are errors, bounded between 0 and 1, unlike WER.
→ Speech recognition evaluation beyond WER - Matrix language
- The language that supplies the grammatical frame of a mixed sentence, into which words from another language are inserted.
→ Code-switching in speech recognition - Mean opinion score (MOS)
- The average of listener ratings on a 1 to 5 scale, used to judge the perceived quality of synthetic or processed speech.
→ What is text-to-speech? - Meaning-preserving edit
- A difference between reference and hypothesis, such as a filler word or a spelling variant, that leaves the intended message unchanged.
→ Semantic word error rate - Media fork
- A copy of a live call's audio stream sent to another service, such as a recogniser, without interrupting the call.
→ Telephony transcription - Media stream
- A provider feature that forks or relays a call's audio to your server, typically as base64-encoded frames over a WebSocket.
→ Voice agents on the phone - MiFID II call recording
- The obligation under Article 16(7) of Directive 2014/65/EU for investment firms to record telephone conversations and electronic communications relating to transactions.
→ Speech recognition in financial services - Minimum charge
- The smallest amount billed per request, regardless of how short the audio is.
→ Speech-to-text pricing models - Morpheme
- The smallest unit of meaning in a word, such as a stem, a prefix or an inflectional ending.
→ Compound words and morphology - Multilingual ASR
- A single speech recognition model trained to transcribe more than one language.
→ Multilingual speech recognition - Multilingual voice agent
- A voice agent that can hold conversations in more than one language, choosing or switching the language of recognition, reasoning and speech per call or per turn.
→ Multilingual voice agents - Multimodal fusion
- Combining predictions or features from text and audio to produce a single classification.
→ Sentiment analysis on speech
N
- Narrowband audio
- Audio sampled at 8 kHz, carrying frequencies up to about 3.4 kHz, as used by the traditional telephone network.
→ Telephony transcription - Natural language understanding (NLU)
- Extracting intent and structured information, such as dates or account numbers, from a user's words.
→ Voice AI vs speech AI vs conversational AI - NER model
- A method for scoring live subtitle accuracy that counts edition and recognition errors, weighted by severity, against the number of words.
→ Media and broadcast transcription - Normalised cross entropy (NCE)
- A confidence metric used in NIST evaluations that measures how much information the scores add over always predicting the base rate of correct words.
→ Confidence scores in speech recognition - Numeric exact match
- The share of numbers in a test set whose written value is transcribed exactly right, regardless of the surrounding words.
→ Numbers and alphanumerics - Nyquist frequency
- Half the sample rate; the highest frequency a sampled signal can represent without aliasing.
→ Audio formats for speech recognition
O
- On-device ASR
- Speech recognition that runs entirely on the user's phone, computer, car or embedded hardware, without sending audio to a server.
→ On-device vs cloud speech recognition - On-premise deployment
- Running software on hardware in the customer's own data centre or colocation space rather than on the vendor's infrastructure.
→ On-premise and VPC deployment - Open ASR Leaderboard
- A public Hugging Face leaderboard that ranks speech recognition models by word error rate on a fixed set of English benchmarks, alongside inverse real-time factor.
→ ASR benchmarks and leaderboards - Open-weight model
- A model whose trained parameters are published so that anyone can run it on their own infrastructure, whether or not its training data and code are also released.
→ Sovereign AI in Europe - Operational sovereignty
- Control over who operates a system and can access its data, independent of where the underlying technology was developed.
→ Sovereign AI in Europe - Orchestrator
- The process that moves audio, text and control events between the components of a voice agent and decides when each one runs.
→ Voice agent architecture - Out-of-vocabulary (OOV) word
- A word that a system's fixed vocabulary does not contain and therefore cannot output as a whole.
→ Compound words and morphology - Over-biasing
- A failure of context biasing in which boosted words or formats are output where they were not spoken, raising the error rate on ordinary speech.
→ Over-biasing - Overlapped speech
- Stretches of audio where two or more people talk at the same time.
→ What is speaker diarization?
P
- Paired bootstrap
- A resampling method that estimates the uncertainty of the difference between two systems scored on the same utterances.
→ Building your own ASR test set - Parameter count
- The number of learned weights in a model; the main driver of its memory footprint and compute cost.
→ Model size and quantization - Parameter-efficient fine-tuning
- Fine-tuning a small number of added or selected parameters, such as adapters or LoRA matrices, while freezing the rest of the model.
→ Fine-tuning vs context biasing - Partial result
- A provisional transcript of the audio received so far, which may be revised as more audio arrives.
→ Streaming speech recognition→ Partial vs final results - pass^k
- A reliability measure: the probability that an agent succeeds on the same task in all of k independent attempts.
→ Testing voice agents - Pause and resume
- Stopping call recording while a customer gives card details, so that sensitive payment data never reaches the recording or transcript.
→ Contact centre speech analytics - Peaky behaviour
- The tendency of CTC models to emit each token in a single frame surrounded by blanks, which makes word boundaries imprecise.
→ Word timestamps and forced alignment - Perplexity
- A measure of how well a language model predicts text; lower perplexity means the model finds the text less surprising.
→ Acoustic models and language models - Personal data
- Any information relating to an identified or identifiable natural person (GDPR Article 4(1)).
→ GDPR and voice data - Phonotactics
- The rules governing which sound sequences occur in a language, used by early LID systems as a statistical fingerprint.
→ Spoken language identification - PII redaction
- Detecting personal data such as names, card numbers and addresses in a transcript or recording and removing or masking it.
→ PII redaction in transcripts - Pilot
- A time-boxed production trial in which a buyer runs a shortlisted vendor on live traffic against success criteria agreed in advance.
→ How to evaluate a speech-to-text vendor - Pitch accent
- A word-level pitch pattern that can distinguish words; Swedish has two, usually called accent 1 and accent 2.
→ Swedish speech recognition - Polyphonic detection
- Detecting several sound events that overlap in time, such as speech over music.
→ Audio event detection - Pooled error rate
- An error rate computed by summing errors and reference words over a whole corpus before dividing, rather than averaging per-utterance rates.
→ Semantic word error rate - Pop-on and roll-up captions
- Two caption display styles: pop-on shows complete blocks at once, roll-up adds words line by line as they are recognised.
→ Media and broadcast transcription - Post-training quantisation (PTQ)
- Quantising an already trained model using a small calibration set, without further training.
→ Model size and quantization - Pre-roll
- Audio kept from just before detected speech onset and prepended to the segment, so the first sound of a word is not lost.
→ What is voice activity detection (VAD)? - Premature cut-off
- An end-of-turn error where the system responds while the user still intended to continue speaking.
→ End-of-turn detection - Processor
- An organisation that processes personal data on behalf of a controller and only on its instructions (GDPR Article 4(8)).
→ Data processing agreements - Prompt injection
- Instructions hidden in input data that a language model mistakes for instructions from its operator.
→ Using transcripts with LLMs - Pronunciation lexicon
- A dictionary mapping each word to one or more sequences of phones, used by classical recognisers to connect acoustic and language models.
→ Acoustic models and language models - Proper noun
- A name of a specific person, place, organisation or product, whose spelling usually cannot be derived from its sound.
→ Rare words and proper nouns - Prosody
- The rhythm, pitch, loudness and timing of speech, which carry meaning beyond the words themselves.
→ Sentiment analysis on speech - Pseudonymisation
- Processing personal data so it can no longer be attributed to a person without additional information kept separately (GDPR Article 4(5)).
→ PII redaction in transcripts - Public task
- The GDPR lawful basis in Article 6(1)(e) for processing necessary to perform a task in the public interest or in the exercise of official authority.
→ Speech recognition in the public sector
Q
- Quantisation
- Storing and computing a model's weights, and sometimes activations, in lower-precision number formats such as 8-bit integers.
→ Model size and quantization
R
- Raddoppiamento sintattico
- In central and southern Italian, the doubling of a word-initial consonant after certain words, so that a casa is pronounced with a long /k/.
→ Italian speech recognition - Read-back confirmation
- The agent repeating a captured value, such as a date or account number, so the caller can confirm or correct it before an action runs.
→ Tool calling in voice agents - Real-time (streaming) transcription
- Transcription that processes audio as it is being spoken and returns text within a fraction of a second.
→ Real-time vs batch transcription - Real-time factor (RTF)
- Processing time divided by audio duration; below 1 means the system is faster than real time.
→ Real-time factor (RTF and RTFx) - Redaction recall
- The share of true personal data items in a transcript that a redaction system finds and removes.
→ PII redaction in transcripts - Reference transcript
- A human-verified transcript treated as ground truth when scoring a speech recognition system.
→ What is word error rate (WER)? - Regional Italian
- Standard Italian as spoken with the accent, intonation and some vocabulary of a particular region, distinct from the local dialect.
→ Italian speech recognition - Repetition loop
- A decoding failure in which an autoregressive model emits the same word or phrase over and over.
→ Hallucinations in speech recognition - Respeaking
- A live subtitling method in which a trained person repeats and punctuates the programme audio into a speech recogniser tuned to their voice.
→ Media and broadcast transcription - Restricted transfer
- Making personal data available to a recipient in a country outside the European Economic Area, which GDPR Chapter V permits only under specific mechanisms.
→ EU data residency for speech AI - Retrieval-augmented generation (RAG)
- Answering a question by first retrieving relevant passages from a store, then passing them to a language model as context.
→ Using transcripts with LLMs - Retroflex sandhi
- In Central Swedish, the merging of r with a following t, d, n, s or l into a single retroflex consonant, including across word boundaries.
→ Swedish speech recognition - Reverberation
- Delayed reflections of sound from walls and surfaces that smear speech over time.
→ Transcribing noisy audio - RNN-T (transducer)
- An architecture that combines an audio encoder with a prediction network over previously emitted tokens, joined to decide at each step whether to emit a token or move to the next frame.
→ ASR model architectures - RTFx
- Inverse real-time factor: seconds of audio processed per second of compute, so higher is faster.
→ ASR benchmarks and leaderboards - RTP
- The Real-time Transport Protocol, which carries audio packets in VoIP and SIP calls.
→ Telephony transcription
S
- Sample rate
- The number of amplitude measurements taken per second of audio, which sets the highest frequency the recording can represent.
→ Audio formats for speech recognition - Segmentation error
- An error in where word boundaries fall, such as splitting a compound into two words, rather than in which sounds were recognised.
→ Character error rate (CER) - Self-supervised learning
- Training a model on unlabelled data by having it predict hidden parts of its own input, such as masked stretches of audio.
→ A brief history of speech recognition - Semantic word error rate
- A variant of word error rate that counts only the edits that change the meaning of a transcript, ignoring differences a reader would treat as equivalent.
→ Semantic word error rate - Sentiment analysis
- Classifying the attitude expressed in language, usually as positive, negative or neutral, sometimes with a strength score.
→ Sentiment analysis on speech - Separable verb
- A German verb whose prefix detaches and moves to the end of the clause, as in anrufen becoming ich rufe dich an.
→ German speech recognition - Service plugin
- An adapter that connects a framework to a specific STT, LLM or TTS provider behind a common interface.
→ Voice agent frameworks - Seseo
- Pronouncing the letters c (before e or i) and z as /s/, the norm in Latin America and parts of southern Spain.
→ Spanish speech recognition - Session
- One streaming connection to a speech-to-text service, from opening the socket and sending configuration to closing it after the last result.
→ Integrating a speech-to-text API - Session affinity
- Routing every message of a stream to the same server, so its state stays in one place.
→ Concurrency and scaling - Shallow fusion
- Adding a weighted external language model score to an end-to-end model's scores during beam search.
→ Acoustic models and language models→ What is domain adaptation in ASR?→ Text-only domain adaptation - Short-lived token
- A temporary credential minted by your server so that a browser or mobile client can connect without holding the long-lived API key.
→ Integrating a speech-to-text API - Signal-to-noise ratio (SNR)
- The power of the speech signal relative to the background noise, expressed in decibels; lower values mean noisier audio.
→ Transcribing noisy audio - Silence timeout
- An endpointing rule that declares the utterance over after a fixed duration of detected silence.
→ What is endpoint detection?→ End-of-turn detection - Silent agreement
- Grammatical endings that are written but not pronounced, such as the plural -ent in ils chantent, which an ASR system must infer from context.
→ French speech recognition - Simulated caller
- An automated counterpart, usually a language model with a persona and goal plus a text-to-speech voice, that holds test conversations with a voice agent.
→ Testing voice agents - SIP trunk
- A connection between a telephony provider and your system that carries calls using the Session Initiation Protocol for signalling and RTP for media.
→ Voice agents on the phone - Slice evaluation
- Reporting error rates separately for subgroups of a test set, such as accent groups, to expose gaps an average hides.
→ Accents and regional speech - SOAP note
- A clinical note structured as Subjective, Objective, Assessment and Plan.
→ Ambient clinical documentation - Sound event detection (SED)
- Labelling which sound classes occur in a recording and their start and end times.
→ Audio event detection - Sovereign AI
- The ability of a country, region or organisation to develop, run and control AI systems under its own laws and without depending on decisions made elsewhere.
→ Sovereign AI in Europe - Speaker diarization
- Partitioning an audio recording into segments labelled by speaker, answering who spoke when without knowing who the speakers are.
→ What is speaker diarization? - Speaker embedding
- A fixed-length vector computed from a stretch of speech that captures the speaker's vocal characteristics, such as an x-vector.
→ Speech recognition vs voice recognition→ What is speaker diarization? - Speaker identification
- Deciding which of a set of enrolled people produced a given stretch of speech.
→ Diarization vs identification vs verification - Speaker recognition
- Determining who is speaking from the characteristics of their voice; the technical name for what products call voice recognition.
→ Speech recognition vs voice recognition - Speaker verification
- Checking whether a voice matches one claimed identity, a yes or no decision.
→ Speech recognition vs voice recognition→ Diarization vs identification vs verification→ Voice biometrics and consent - Speaker-disjoint split
- A division of a corpus in which no speaker in the test split appears in the training split.
→ Benchmark contamination - Speech AI
- The family of technologies that process or generate spoken language: recognition, synthesis, speaker recognition and related audio analysis.
→ Voice AI vs speech AI vs conversational AI - Speech analytics
- Transcribing recorded or live calls and analysing the text to measure topics, compliance, quality and customer outcomes.
→ Contact centre speech analytics - Speech emotion recognition (SER)
- Predicting an emotional state, such as anger or sadness, from acoustic properties of the voice.
→ Sentiment analysis on speech - Speech-to-speech model
- A single model that consumes audio and generates audio directly, without an explicit text transcript in between.
→ What is voice AI?→ Speech-to-speech models vs pipelines - Spoken Dutch Corpus (CGN)
- A reference corpus of around nine million words of contemporary Dutch speech from the Netherlands and Flanders, across many speaking styles.
→ Dutch speech recognition - Spoken form
- Text written exactly as it is pronounced, with numbers, symbols and abbreviations spelled out as words.
→ Punctuation, casing, and inverse text normalization→ Numbers and alphanumerics - Spoken language identification (LID)
- Automatically determining which language is spoken in a stretch of audio.
→ Spoken language identification - SRT (SubRip)
- A plain-text subtitle format of numbered cues with comma-separated millisecond timestamps.
→ Captions and subtitles: SRT, VTT, and timing - Stability
- How rarely the words in partial results are revised before they become final.
→ Partial vs final results - Standard contractual clauses (SCCs)
- Model contract terms adopted by the European Commission that provide appropriate safeguards for transfers of personal data outside the EEA.
→ EU data residency for speech AI - Statement of Applicability (SoA)
- The ISO 27001 document listing which Annex A controls an organisation applies, which it excludes, and why.
→ ISO 27001 for speech AI vendors - Stitching
- Merging the transcripts of adjacent segments into one continuous transcript, resolving any overlap between them.
→ Long-form transcription - Storage limitation
- The GDPR principle, in Article 5(1)(e), that personal data is kept in identifiable form no longer than necessary for its purpose.
→ Zero data retention - Stratified evaluation
- Reporting a metric separately for subgroups of the test set, such as accent, device or noise level, rather than only as one average.
→ Speech recognition evaluation beyond WER - Stratified sampling
- Selecting test clips so that each relevant condition, such as channel, language or accent, is represented in a known proportion.
→ Building your own ASR test set - Streaming speech recognition
- Recognition that consumes audio incrementally and emits text while the speaker is still talking.
→ Streaming speech recognition - Structured identifier
- A code with a fixed format, such as an order number, licence plate, IBAN or SKU, whose characters carry no linguistic meaning.
→ Transcribing structured identifiers - Structured output
- Model output constrained to a schema, such as JSON with named fields, so it can be validated and parsed.
→ Using transcripts with LLMs - Sub-processor
- A third party that a vendor engages to process personal data on a customer's behalf, such as a cloud provider.
→ How to evaluate a speech-to-text vendor→ Data processing agreements - Sub-word token
- A unit of text between a character and a word, such as 'milli' or 'grams', learned from data so that any word can be spelled from a fixed vocabulary.
→ How speech-to-text works - Subtitles for the deaf and hard of hearing (SDH)
- Captions that include speaker identification and relevant non-speech sounds as well as dialogue.
→ Captions and subtitles: SRT, VTT, and timing - Subword tokenisation
- Splitting text into units smaller than words, learned from data with algorithms such as byte-pair encoding or a unigram language model, so that a model can build rare words from common pieces.
→ Compound words and morphology - Subword unit
- A fragment of a word, learned from data, that end-to-end recognisers use so they can spell words never seen whole in training.
→ Rare words and proper nouns - Särskrivning
- The Swedish term for writing a compound as separate words, which can change its meaning.
→ Swedish speech recognition
T
- Tail latency
- Latency at high percentiles such as p95 or p99, which describes the slowest responses users experience.
→ Speech-to-text latency→ Concurrency and scaling - Task completion rate
- The share of test or production conversations in which the caller's goal was achieved correctly.
→ Testing voice agents - Test split
- The portion of a dataset reserved for evaluation, which should never be seen by a model during training or tuning.
→ ASR benchmarks and leaderboards - Text normalisation (scoring)
- Rewriting reference and hypothesis into a common form (case, punctuation, numbers) before computing an error rate.
→ What is word error rate (WER)? - Text-only domain adaptation
- Improving a speech recognition model on a specialist domain using only written text from that domain, with no transcribed audio.
→ Text-only domain adaptation - Text-to-speech (TTS)
- Software that converts written text into synthetic spoken audio.
→ What is text-to-speech? - Throughput
- The total amount of audio a system processes per unit of time, across all concurrent requests.
→ Real-time factor (RTF and RTFx) - Time to final transcript
- The delay between the end of speech and the moment a streaming system returns the final transcript for it.
→ How to evaluate a speech-to-text vendor - Time to first audio
- The delay between sending text to a TTS system and receiving the first playable chunk of audio.
→ What is text-to-speech? - Time to first partial
- The delay between the start of speech and the first interim transcript the client receives.
→ Speech-to-text latency - Time to first sound
- The delay between the caller finishing speaking and the agent's first audible response.
→ What is a voice agent? - Time to first token
- The delay between the start of speech and the first partial word reaching the client.
→ Partial vs final results→ Voice agent latency budget - Timestamp drift
- A growing offset between transcript timestamps and the true position of words in the audio over a long recording.
→ Long-form transcription - Tool calling
- A language model emitting a structured request, with a function name and arguments, for the application to execute and return a result.
→ Tool calling in voice agents - Tool schema
- A machine-readable description of a tool's name, purpose and parameters, usually in JSON Schema, supplied to the model with the prompt.
→ Tool calling in voice agents - Trader turret
- A specialised multi-line telephone system used on trading floors, often with many open lines and speaker channels.
→ Speech recognition in financial services - Transcription guidelines
- A written specification of how reference transcribers handle fillers, numbers, names, overlap and unclear speech.
→ Building your own ASR test set - Transition relevance place
- A point in conversation where a change of speaker would be appropriate, as described in conversation analysis.
→ VAD vs endpointing vs turn detection - Transition-relevance place
- A point in speech, identified in conversation analysis, where a change of speaker becomes appropriate.
→ End-of-turn detection - Transport
- The layer that carries audio between the user and the agent, such as WebRTC, a WebSocket or a SIP telephone call.
→ Voice agent architecture - TTFS
- The time from the end of a user's speech to the delivery of the final transcript for that turn.
→ Voice agent latency budget - TTS time to first byte
- The delay between sending text to a speech synthesiser and receiving the first chunk of playable audio.
→ Voice agent latency budget - Turn detection
- Deciding whether a speaker in a conversation has finished their turn and yielded the floor.
→ VAD vs endpointing vs turn detection - Tussentaal
- The informal spoken register in Flanders between local dialect and standard Dutch.
→ Dutch speech recognition - Two-pass transcription
- Running a fast streaming pass for live use and a second, full-context pass on the finished recording for the stored transcript.
→ Real-time vs batch transcription
U
- Unicode normalisation
- Converting text to a canonical Unicode form, such as NFC, so that visually identical characters are stored as the same code points before comparison.
→ Character error rate (CER) - Utterance
- A continuous stretch of speech, usually bounded by pauses, that a recogniser processes as one unit.
→ What is speech recognition?
V
- Verbatim report
- The official record of parliamentary debates, edited from what was said to remove repetition and obvious slips while keeping the meaning, such as Hansard in the UK.
→ Speech recognition in the public sector - Verbatim transcript
- A transcript that records every word as spoken, including false starts, repetitions and fillers.
→ Legal transcription - Virtual private cloud (VPC)
- A logically isolated network inside a public cloud account, controlled by the account owner, in which vendor software can be deployed.
→ On-premise and VPC deployment - Vocoder
- The model that turns an intermediate acoustic representation, such as a mel spectrogram, into an audio waveform.
→ What is text-to-speech? - Voice activity detection (VAD)
- Classifying short frames of audio as speech or non-speech.
→ What is voice activity detection (VAD)? - Voice agent
- Software that holds a spoken conversation with a person in real time, understanding what they say and responding in synthesised speech, usually able to take actions on their behalf.
→ What is a voice agent? - Voice agent framework
- A software library that provides the real-time orchestration layer of a voice agent: transports, streaming pipelines, turn handling and integrations with speech and language services.
→ Voice agent frameworks - Voice AI
- Software that takes spoken audio as input, produces spoken audio as output, or both, using machine learning models.
→ What is voice AI? - Voice biometrics
- Recognising who is speaking from the characteristics of their voice, for verification or identification.
→ Voice biometrics and consent - Voice surveillance
- Monitoring recorded trader and client calls to detect market abuse, misconduct or rule breaches.
→ Speech recognition in financial services - Voice-to-voice latency
- The time from the moment a user stops speaking to the moment they hear the first audio of the agent's reply.
→ Voice agent architecture - Voiceprint
- A stored representation of a person's voice, usually a speaker embedding, used as the reference for later comparisons.
→ Voice biometrics and consent - Voseo
- Using vos with its own verb forms instead of tú for the informal second person, as in Argentine vos tenés.
→ Spanish speech recognition - Vowel reduction
- Weakening or deleting unstressed vowels; in European Portuguese it often removes whole syllables from the audio.
→ Portuguese speech recognition
W
- Wake word
- A keyword that activates a device or assistant, after which fuller speech recognition begins.
→ Keyword spotting and wake words - Weak labels
- Training labels that say a sound is present in a clip but not where, as in AudioSet.
→ Audio event detection - WebVTT
- The W3C caption format for HTML video, similar to SRT but with a header, dot-separated milliseconds and support for positioning and styling.
→ Captions and subtitles: SRT, VTT, and timing - Weighted finite-state transducer (WFST)
- A rule-based automaton that maps input strings to output strings, widely used to implement normalisation grammars.
→ Punctuation, casing, and inverse text normalization - Word error rate (WER)
- The number of word substitutions, deletions and insertions needed to turn a transcript into its reference, divided by the number of words in the reference.
→ What is word error rate (WER)? - Word timestamp
- The start and end time of a recognised word within the audio.
→ Word timestamps and forced alignment - Written form
- Text in the conventions a reader expects, with digits, currency symbols, dates and abbreviations.
→ Punctuation, casing, and inverse text normalization→ Numbers and alphanumerics
Y
- Yeísmo
- The merger of the sounds written ll and y, so that calló and cayó sound identical.
→ Spanish speech recognition
Z
- Zero data retention (ZDR)
- A processing model in which a provider handles customer content only for the duration of a request and does not persist it afterwards.
→ Zero data retention - Zero-shot evaluation
- Evaluating a model on a dataset whose training split it has not used, so the score reflects generalisation to a new distribution.
→ Benchmark contamination