Glossary
The Voice AI
Glossary
Plain definitions of the terms teams use to build, test, and monitor voice agents.
- Terms
- 149
- Last updated
- October 29, 2026
A(7)
- Agent Behavior CloningReplicating human conversational patterns to train more natural-sounding agents.
- Agent PersonaThe personality and communication style of the voice agent.
- Agent ProfileConfiguration settings that define an AI agent's personality, behavior, and capabilities.
- Agentic AIAI systems designed to autonomously pursue complex goals with limited direct human supervision.
- Audio CodecAlgorithm for compressing and decompressing audio data.
- Automatic Speech Recognition (ASR)Technology that converts a caller's spoken words into text that voice agents can understand and process in real-time.
- Average Handle Time (AHT)Average duration of a call from start to finish.
B(2)
C(25)
- Call Abandonment RatePercentage of callers who hang up before their issue is resolved.
- Call AnalyticsAnalysis of voice conversation data to extract insights about performance, sentiment, and outcomes.
- Call EscalationProcess of transferring callers from voice agents to human agents when issues exceed AI capabilities.
- Call MonitoringReal-time observation and analysis of ongoing calls.
- Call QueuingSystem for managing multiple incoming calls and wait times.
- Call RecordingCapturing and storing audio from voice conversations for analysis or compliance.
- Call RoutingProcess of directing incoming calls to appropriate destinations based on context and rules.
- Channel DiarizationSeparating speakers in multi-channel audio where each channel represents a different speaker.
- Cold Start LatencyDelay when initializing a voice agent system from an idle state.
- Complexity ThresholdPoint at which a task exceeds AI capabilities and requires human intervention.
- Concurrent CallsNumber of simultaneous voice conversations an AI system can handle at once.
- Confidence ScoreNumerical measure of how certain an AI system is about its transcription or decision.
- Context WindowThe amount of conversation history the model can consider.
- Contextual GroundingTechnique ensuring AI responses are based on verified, relevant source information rather than hallucinated content.
- Conversation AnalyticsTools that analyze voice agent calls to measure performance, identify issues, and understand caller satisfaction.
- Conversation BufferingManaging conversation history within context window limitations.
- Conversation IntelligenceVoice agent's ability to understand context, emotion, and nuance in caller speech to provide intelligent responses.
- Conversation MemoryAI system's ability to remember and reference earlier parts of the same conversation.
- Conversation State MachinesStructured models that define conversation states, transitions, and fallbacks.
- Conversation TreesVisual representation of all possible conversation paths and branches.
- Conversational AIAI systems that engage in natural spoken dialogue, powering voice agents to handle phone conversations like humans.
- Conversational AI BenchmarkingStandardized testing to measure and compare agent performance across metrics.
- Conversational AI Compliance ToolsTools ensuring voice agents meet regulatory and ethical requirements.
- Conversational PathwaysStructured conversation flows that guide AI agent responses based on user input and business logic.
- Custom LLM IntegrationConnecting proprietary or self-hosted language models to voice platforms.
D(3)
- Dialogue FlowThe designed conversation path voice agents follow, including branches for different caller responses and scenarios.
- DTMF (Dual-Tone Multi-Frequency)Touch-tone signals sent when callers press phone keypad buttons, allowing voice agents to receive numeric input.
- Dynamic RoutingReal-time decision making about where to send calls based on context, sentiment, and availability.
E(7)
- Echo CancellationTechnology that removes audio feedback and echo from calls.
- Edge Voice ProcessingRunning voice AI computations on local devices near users.
- EmbeddingsNumerical vector representations of text or audio that capture semantic meaning.
- End-to-End EncryptionSecurity method where data is encrypted from sender to receiver with no intermediate decryption.
- End-to-End LatencyTotal time from when a caller stops speaking to when they hear the voice agent's response, including all processing steps.
- EndpointingDetecting when a user has finished speaking to begin processing.
- Entity ExtractionIdentifying and extracting specific data points from user speech.
F(5)
- Fallback ResponsePredefined phrases voice agents use when they don't understand callers or encounter system errors.
- Few-shot LearningAI capability to learn new tasks from just a few examples.
- Fine-tuningProcess of training an AI model on specific data to improve performance for particular tasks.
- First Call Resolution (FCR)Metric measuring percentage of customer issues resolved in a single interaction.
- Function CallingThe ability for voice agents to trigger external APIs or actions during calls, like booking appointments or checking inventory.
G(4)
- GDPR ComplianceEuropean privacy regulation governing collection and processing of personal data.
- Golden DatasetsCurated high-quality conversation examples representing ideal agent behavior.
- Group Chat OrchestrationManaging multi-agent conversations where agents collaborate on complex tasks.
- GuardrailsSafety boundaries ensuring voice agents stay on-brand, don't make unauthorized promises, and handle sensitive information properly.
H(2)
I(3)
- Intent CoveragePercentage of user intents that an AI agent can successfully identify and handle.
- Intent RecognitionThe voice agent's ability to understand what a caller wants to accomplish from their spoken words.
- IVR (Interactive Voice Response)Traditional phone menu systems ("Press 1 for sales") that voice agents are increasingly replacing with natural conversation.
J(1)
L(4)
- Language ModelThe AI brain powering voice agents, determining how they understand callers and generate appropriate spoken responses.
- LatencyThe delay between when a user speaks into a voice agent and when they hear the agent's spoken response, measuring the complete round-trip time from voice input to voice output.
- LiveKit Agents FrameworkFramework for building and deploying scalable voice agents with automatic load balancing.
- LLM Voice CachingStoring frequently used responses to reduce latency and costs.
M(8)
- Media StreamContinuous flow of audio/video data in real-time communication.
- Memory Management HierarchiesTiered system for managing different types of agent memory and context.
- Model Context Protocol (MCP)Standardized way for AI agents to securely connect to external tools and data sources.
- Multi-Agent HandoffTransferring conversations between specialized agents based on context.
- Multi-modal Voice InteractionsIntegrating voice with visual, text, and haptic feedback for richer experiences.
- Multi-turn ConversationDialog involving multiple back-and-forth exchanges rather than single question-answer pairs.
- Multilingual SupportAbility to handle conversations in multiple languages.
- Multimodal AIAI systems that can process and respond to multiple input types like text, voice, and images.
N(4)
- Natural Language Processing (NLP)AI technology that helps machines understand, interpret, and generate human language.
- Natural Language Understanding (NLU)Voice agent's ability to comprehend the meaning behind spoken words, not just transcribe them.
- Noise ReductionTechnology that filters background noise from audio to improve speech recognition accuracy.
- Noise SuppressionAlgorithms that remove background noise from audio signals.
O(2)
P(13)
- p50 LatencyThe median latency: 50% of requests are faster than this value.
- p90 Latency90th percentile latency: 90% of requests are faster than this value.
- p99 Latency99th percentile latency: 99% of voice agent responses are faster than this value, representing the worst 1% of experiences.
- Packet LossNetwork packets that fail to reach their destination.
- PCI ComplianceSecurity standards for handling credit card information in voice systems.
- PII RedactionAutomatically removing personally identifiable information from transcripts and audio.
- Polyphonic AISystem where specialized AI agents with different expertise work together on complex tasks.
- Production Voice MonitoringReal-time tracking of voice agent performance in live environments.
- Profanity FilteringFeature that detects and handles inappropriate language in speech recognition results.
- Prompt EngineeringThe practice of crafting instructions that control how voice agents speak, behave, and handle different caller scenarios.
- Prompt TemplateReusable structure for creating consistent AI agent prompts.
- Pronunciation AssessmentAI evaluation of speech pronunciation quality with accuracy, fluency, and completeness scores.
- ProsodyThe rhythm, stress, and intonation patterns in synthesized speech.
Q(1)
R(4)
- RAG (Retrieval-Augmented Generation)Technique that enhances AI responses by retrieving relevant information from a knowledge base before generating answers.
- Rate LimitingRestricting the number of API calls or requests within a time period.
- Real-time ProcessingProcessing audio or generating responses immediately as data is received.
- Real-time TranscriptionConverting caller speech to text instantly as they talk, enabling voice agents to process and respond quickly.
S(13)
- Semantic AccuracyMeasure of how well AI understands the meaning and intent behind user requests.
- Semantic SearchSearch method that finds content based on meaning rather than exact keyword matches.
- Sentiment AnalysisProcess of detecting positive, neutral, or negative emotional tone in a speaker's voice or language.
- Session StateThe maintained data and context throughout a single conversation session.
- Silence DetectionIdentifying periods of no speech in audio streams.
- SIP (Session Initiation Protocol)Protocol for initiating, maintaining, and terminating voice calls over IP networks.
- SIP TrunkingMethod of connecting telephony systems to AI voice agents using Session Initiation Protocol.
- SOC 2 Type IISecurity audit standard verifying controls for data protection over extended time period.
- Speaker DiarizationProcess of segmenting audio recordings to identify 'who spoke when' by separating individual speakers.
- State TransitionsThe movement between different conversational states in multi-agent voice systems, where specialized agents handle specific parts of a call journey.
- StreamingProcess of generating and transmitting audio/text in real-time chunks rather than waiting for complete generation.
- Streaming ProcessingVoice agents processing audio continuously as callers speak, rather than waiting for complete sentences, enabling faster responses.
- STUN ServerSession Traversal Utilities for NAT, a protocol that helps establish peer-to-peer connections through firewalls.
T(13)
- TemperatureParameter controlling randomness in AI text generation, from deterministic (0) to creative (1).
- Test CaseA specific scenario designed to validate voice agent behavior, including the caller's input, expected agent responses, and success criteria.
- Text NormalizationProcess of converting spoken forms like 'twenty dollars' to written forms like '$20'.
- Text-to-Speech (TTS)Technology that converts a voice agent's text responses into natural-sounding synthesized speech that callers hear.
- Time to First Word (TTFW)Time elapsed from when the user stops speaking to when the voice agent begins playing its first audible word in response.
- Time-to-First-Audio (TTFA)Time from user input completion to when synthesized speech begins playing.
- Time-to-First-Token (TTFT)Metric measuring how quickly an LLM begins generating its response after receiving input.
- Tool UseVoice agent's ability to access external tools and APIs during calls to retrieve information or perform actions.
- Transcription AccuracyMeasure of how correctly speech recognition converts audio to text, typically expressed as percentage.
- Transcription ConfidenceScore indicating the ASR system's certainty about transcription accuracy.
- Transient AssistantsDynamically generated agents created for each conversation based on context.
- Turn DetectionProcess of determining when a user begins or ends their conversational turn, deciding when the AI should start listening and when to respond.
- Turn-takingThe coordination of who speaks when in voice agent calls, mimicking natural human dialogue rhythm.
U(1)
V(21)
- Vector StoreDatabase that stores embeddings of knowledge base content for semantic search and retrieval.
- Voice Activity Detection (VAD)Algorithm that determines when a caller starts and stops speaking, enabling voice agents to know when to listen and when to respond.
- Voice Agent A/B TestingComparing different agent versions to determine which performs better.
- Voice Agent Fine-TuningAdjusting pre-trained models on domain-specific data for improved performance.
- Voice Agent Load BalancingDistributing voice calls across multiple agent instances for optimal performance.
- Voice Agent ObservabilityReal-time monitoring of voice agents in production to track performance, errors, and caller experience.
- Voice Agent Quality Index (VAQI)Deepgram's metric combining multiple factors to measure overall voice agent performance quality.
- Voice Agent ReasoningAI's ability to perform logical inference and complex problem-solving during dialogue.
- Voice Agent Regression DetectionAutomatically identifying when agent performance degrades after changes.
- Voice Agent TemplatesPre-built agent configurations for common use cases.
- Voice Agent TestingSystematic validation of voice agents through simulated calls to ensure they handle real conversations correctly.
- Voice AI AgentAn autonomous conversational system that combines speech recognition, natural language processing, and text-to-speech synthesis to enable hands-free, real-time interactions through spoken communication.
- Voice AI Evaluation MetricsComprehensive metrics for assessing voice agent quality beyond basic accuracy.
- Voice AI InfrastructureThe complete technical stack required to run voice AI systems at scale.
- Voice AnalyticsTools that analyze voice agent calls to extract insights about caller behavior, sentiment, and conversation outcomes.
- Voice BiometricsTechnology that identifies individuals based on unique characteristics of their voice.
- Voice CloningTechnology that creates synthetic speech matching a specific person's voice characteristics.
- Voice DesignCreating custom AI voices using text prompts without voice samples.
- Voice IntelligenceAdvanced analytics and insights derived from voice conversations.
- Voice User Interface (VUI)Interface that allows users to interact with systems through voice commands rather than visual elements.
- Voicemail DetectionAI capability to identify when a call reaches an answering machine or voicemail system.
W(5)
- Wake Word DetectionTechnology that activates a voice system when it hears a specific trigger phrase.
- WebhookHTTP callbacks that connect voice agents to your business systems, enabling real-time data exchange during calls.
- WebRTCWeb Real-Time Communication protocol enabling peer-to-peer audio/video communication with minimal latency.
- Word Error Rate (WER)A metric measuring how accurately a voice agent's speech recognition system transcribes spoken words, calculated as errors divided by total words.
- Word-level TimestampsPrecise timing information showing exactly when each word was spoken in an audio recording.